SIMULATION OF SENSORINEURAL HEARING IMPAIRMENT by Paul Duchnowski S.B., Massachusetts Institute of Technology, Cambridge (1987) SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE IN ELECTRICAL ENGINEERING at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY May 1989 Massachusetts Institute of Technology, 1989 Signature redacted Signature of Au thor Department of Electrical Engineering and Computer Science Signature redacted - Certified by Of -- Patrick M. Zurek ignature redactedm i Supervisor - Accepted by v Arthur C. Smith Chairman, Department Committee on Graduate Students JULii1~ A R~ I I V'-s Simulation of Sensorineural Hearing Impairment by Paul Duchnowski Submitted to the Department of Electrical Engineering and Computer Science on May 19, 1989 in partial fulfillment of the requirements for the degree of Master of Science in Electrical Engineering Abstract An algorithm to simulate the effects of sensorineural hearing impairment is investigated. It employs automatic gain control in independent frequency bands to reproduce elevated audibility thresholds, loudness recruitment and reduced dynamic range characteristic of this type of loss. The performance of three listeners with severe sensorineural hearing loss on several speech intelligibility tests is compared to that of normal subjects listening to the output of the simulation. These tests include consonant-vowel syllable identification and sentence keyword identification for several combinations of speech-to-noise ratio, frequency-gain characteristic and overall level. Generally, the simulation algorithm reproduces intelligibility performance very well. There is a clear trend for the simulation to give better intelligibility than observed for impaired listeners when a high-frequency emphasis places more speech above threshold at higher frequencies. It is speculated that upward spread of masking which is not accounted for in the algorithm is the suprathreshold effect responsible for this discrepancy. It is also suggested that the general form of the algorithm may be expanded to accomodate this effect. Thesis Supervisor: Title: Patrick M. Zurek Principal Research Scientist 3 Acknowledgements So, as has been pointed out to me, this is one place in this document where I exercise full editorial conrol. Wow! Power! Kidding aside (for the moment), I would like to thank all of the members of the Sensory Communications Group (a.k.a. the seventh floor gang) for their invaluable assistance and advice throughout this project; they made it a much enriching and enjoyable experience. Special thanks go to Pat Zurek for the insight and wisdom he shared with me and for the patience, above and beyond the call of duty, which he showed when the going was rough. I tip my hat to Julie Greenberg and Doug Henderson, fellow thesis-finishers who shared their Latex, Macintosh and Graf skills: looks like we're going to make it. Please forgive my apocalyptic pronouncements and less than inspired renditions of Pink Floyd's greatest hits during the final hours. I thank my subjects: PG, AL, PB, KG, DW and AW for their good humor in face of lonely hours in front of a terminal inside a sound proof booth. They managed to make it fun. My parents were as usual a constant source of support and encouragement, without them I would not be writing these pages. I can only hope to some day return the love and friendship they showed. I thank Suzy who did what friends do best: patiently listened to my griping and then convinced me that nothing was my fault and that everything was going to be all right. Finally, I gratefully acknowledge the financial support of the General Electric Foundation and the National Institutes of Health who at various stages funded my studies. This thesis is dedicated to my brother Michael, my truest friend always. Contents Abstract Introduction Motivation 1.2 Background . . . . . . 1.3 . 7 7 . Filtering . . . . 1.2.2 Noise masking 1.2.3 Center-Clipping. 1.2.4 Level expansion 8 . 1.2.1 . 8 . 9 9 10 Problem Statement . . Early Algorithms 12 . . 12 . . 13 . . 14 . . . . . . The Center-Clipping Algorithm . . . . . . . . 2.1.1 Rationale . . . . . . . . .. .. . . . 2.1.2 Implementation . . . .. . . . . . . . 2.1.3 Results . . . . . . . . . . . . . . . . . Frequency domain clipping . . . . .. ..... 16 2.2.3 The AGC Algorithm . 2.2.2 Description . . . . . . . . . . . . . . . . . Perceptual effects . . . . . . . . . . . . . . Attempted improvements . . . . . . . . . 17 . 2.2.1 18 . 2.2 12 . 2.1 19 Overview . . . . . . . 3.2 Functional blocks . . . . . 21 3.2.1 The Filterbank . . 21 3.2.2 . 20 3.1 Level detector . . . 24 3.2.3 Gain Computation . 25 3.3 . 20 . . . 3 . . . . . . Implementation . . . . . . 29 . 2 7 1.1 . 1 2 4 CONTENTS 4.1 Aim ...... 35 4.2 Methods ........................... ......................... 35 Speech materials .............. 35 4.2.2 Speech-spectrum noise and frequency-gain characteristics 37 4.2.3 Simulation parameters ........... 38 4.2.4 Subjects . . . . . . . . . . . . . . . . . . 39 4.2.5 Procedure . . . . . . . . . . . . . . . . . 40 R esults . . . . . . . . . . . . . . . . . . . . . . . 42 . . . 4.2.1 49 5.1 A im . . . . . . . . . . . . . . . . . . . . . . . . 49 5.2 Methods...... 49 . Detailed Evaluation of AGC Algorithm ....................... Speech materials . . . . . . . . . . . . . 49 5.2.2 Simulation parameters . . . . . . . . . . 50 5.2.3 Subjects . . . . . . . . . . . . . . . . . . 50 5.2.4 Procedure . . . . . . . . . . . . . . . . . 51 R esults . . . . . . . . . . . . . . . . . . . . . . . 53 5.3.1 CV Syllables . . . . . . . . . . . . . . . 53 5.3.2 Sentences . . . . . . . . . . . . . . . . . 59 5.3.3 Simulation without recruitment . . . . . 67 . . . . . . . . 5.2.1 Discussion 70 6.1 Evaluation of performance 70 6.2 Importance of recruitment 72 6.3 Conclusions 73 . . . . . . . . Simulation Program 77 A.1 Full code . . . . . . . . 77 A.2 Program changes for no recruitment 85 . A 31 35 5.3 6 . . . . . . . . . . . . . . . . . . . . . . Pilot Evaluation of AGC Algorithm 4.3 5 . . . . . .. . 4 Illustration . 3.4 5 CONTENTS B Consonant Confusion Matrices 6 87 Chapter 1 Introduction 1.1 Motivation An accurate simulation of sensorineural hearing loss should prove of value from both a theoretical and practical standpoint. It would allow researchers to study in isolation the perceptual difficulties due to deficiencies in the inner ear. It would separate these problems from those stemming from conductive loss or environmental factors. It would also help to define the aspects of hearing loss that most affect the perception of sound by the hearing impaired. It would provide not only insights into promising compensatory processing but also a preliminary test of the effectiveness of any such method. If speech processed by a proposed hearing aid and subjected to the simulation proved more intelligible than without the compensation, one could consider such a hearing aid as likely to be useful in practice. The value of the simulation would stem from the ease and expediency with which the investigator could use it and thus obtain a quick evaluation of his approach. The simulation could also be used by medical personnel working with the hearing impaired to gain insight regarding the effectiveness of rehabilitative techniques. Finally, it would allow the families and associates of hearing impaired persons to better understand and appreciate, and thus deal with, the impairment. 1.2 Background There have been four types of signal degradation proposed to simulate the effects of hearing impairment. They are described here briefly. 7 CHAPTER 1. INTRODUCTION 1.2.1 8 Filtering In this approach frequency-specific attenuation is used to reproduce the elevated thresholds of the impaired subject for the normal listeners. This is accomplished by a linear filter with an appropriately shaped frequency response. The method does not address any suprathreshold effects such as recruitment. Fabry and Van Tasell (3) used this approach to simulate the impairments of six subjects with losses mainly in the high frequencies and tested the identification of 20 consonant-vowel syllables. There was no significant difference in performance between simulated normal and impaired ears for two of the six impairements subjects when percent correct scores were considered. Four impairments were judged well matched when feature transmission error patterns were compared. Walden et al. (22) performed a similar experiment also using syllables. They tested eight unilaterally impaired subjects and found the patterns of feature recognition similar while mean consonant recognition was generally lower in the impaired ear than in the simulated normal ear. Both sets of authors concluded that threshold shift alone is not sufficient to explain the performance of impaired subjects. They also noted that there appears to be no correlation between gross error scores and feature error patterns. 1.2.2 Noise masking This is perhaps the most popular method which yields reasonable results. It employs additive masking noise which elevates the thresholds of the normal-hearing subjects to match the impairment. The noise also produces abnormally rapid growth of loudness above threshold (recruitment) similar to that seen in sensorineural impairment. In the second part of the above cited study Fabry and Van Tasell used this method to simulate the impairments previously simulated by filtering. Their surprising finding was that noise masking never proved successful when filtering did not. This is in contrast to the consonant reception results of Zurek and Delhorne (24) who also employed noise masking. Their experiments, described in some detail CHAPTER 1. INTRODUCTION 9 in Chapter 4, showed close matching between the performance of impaired and normal subjects for a wide range of impairments and simuli. Whether successful or not, this method suffers from an obvious drawback: the presence of the noise. It precludes simulation of severe or profound losses since the noise required is unacceptably loud. Even at moderate levels subjects may be reluctant to listen to prolonged presentations. The method is effectively limited to losses of less than 60-70 dB HL. Another potential problem lies in the fact that, physiologically, the roots of decreased intelligibility for noise masked normal listeners appear to be different than for the hearing impaired (14). 1.2.3 Center-Clipping Gagne and Erber (4) developed a two-channel device which subjects the input to center-clipping and time/frequency jittering. The center-clipping function (see Figure 2.1), described in detail in the following chapter, simulates the elevated threshold and produces recruitment. The frequency jitter was designed to simulate reduced frequency selectivity. The authors simulated three theoretical audiometric configurations for a series of tests of PB-word recognition and vowel and consonant identification. They reported results in agreement with the published data on similar impairments. They did not, however, conduct comparative studies with actual hearing impaired listeners. As Gagne and Erber themselves point out, there are many ways of distorting speech that will decrease intelligibility. Lower perception scores achieved by simulated normals do not necessarily indicate that they perceive the stimulus just as the impaired listener does. In fact, there appear to exist significant shortcomings in their approach: the system seems to introduce spurious distortion which is unlikely to be a correlate of an impairment (see Chapter 2). 1.2.4 Level expansion Villchur (20; 21) devised a system which adjusts the level of the signal in 16 channels to transpose the impaired subject's loudness relationships to the normal hearing CHAPTER 1. INTRODUCTION 10 span. This transposition was accomplished by projecting, in each channel, the span of hearing of the impaired subject onto the span of hearing of the normal subject. The lower limits were defined as thresholds of audibility. The upper limit for normals was a 74-phon equal loudness contour; for impaired subjects it was a frequency contour about 20-30 dB above the normals' level. The method would thus simulate elevated threshold and recruitment. Four unilaterally impaired subjects judged the effect of the simulation on speech to be "similar" to "very similar" to the unprocessed speech heard in their impaired ears. Villchur also used the algorithm to explain the reported resistance of impaired listeners to white noise and their vulnerability to speech-spectrum noise. The main deficiency of this method is its rather arbitrary definition of recruitment ranges which do not conform results of published studies of loudness growth (6; 12). This problem is illustrated in section 3.2.4. Nonetheless, in the course of this study this approach was judged as the most promising and the algorithm described in Chapter 4 is based on it. 1.3 Problem Statement An issue directly related to the simulation problem is the question of which deficiencies associated with impairment are most closely and directly correlated with reduced speech discrimination ability. Possible candidates include: distortion of the normal loudness relationships (recruitment), reduction in frequency selectivity, greater spread of masking effects, poorer temporal processing, etc. In this study several simulation algorithms were proposed and investigated. Rather than attempting to simulate every possible type of deficiency associated with hearing loss these algorithms sought to reproduce as faithfully as possible a reduced set of these phenomena. This was done for several reasons. First, with more than a few aspects under simulation the problem becomes rather intractable. Second, the extent to which some of these characteristics are present and influence intelligibility is hard to quantify and still a matter of debate. Scharf and Florentine (16), for example, point to decreased frequency resolution as a basic characteristic CHAPTER 1. INTRODUCTION 11 of an impairment. Humes (8), on the other hand, speculates that it may simply be a result of elevated thresholds and injudicious choice of stimulus levels by some investigators. On the other hand, elevated thresholds and recruitment are well accepted (in fact, they are the defining) characteristics of sensorineural hearing loss. Third, one of the goals of this work was precisely to establish the sufficiency of elevated thresholds and recruitment for accurate simulation. By excluding some of the features of loss from the simulation, we are at the outset conceding that some elementary effects (masking, for example) may not be represented. One of the aims, however, was precisely to find whether these features are crucial to explaining reduced speech reception by the hearing impaired. Speech perception was the main focus of this work. The primary motivation for this choice was the simple fact that speech is arguably the most important acoustic stimulus for humans and thus should be our first concern. Chapter 2 Early Algorithms During the course of this study several algorithms which might potentially meet the criteria outlined above were explored. This chapter describes two which were considered but subsequently found to suffer from significant shortcomings. The rationale behind them and the reasons for judging them unsatisfactory will be briefly described here. 2.1 2.1.1 The Center-Clipping Algorithm Rationale This approach was motivated by the system described by Erber and Gagne (4). It is based on the observation that the center-clipping non-linear transfer function shown in Figure 2.1 reproduces both the threshold and recruitment effects associated with hearing loss. Thus, input signals whose amplitude is below the cutoff point are attenuated to zero. For an appropriately chosen cutoff point this function would therefore simulate the elevated threshold. Furthermore, as the input amplitude grows above the cutoff the rms value of the output grows faster than linearly and asymptotically approaches that of the input. In this sense the function mirrors the effects of the abnormally steep loudness growth experienced by impaired listeners. Figure 2.2 shows as an example the effect of the center-clipper on a sine-wave at three different amplitudes. It is evident that without further processing the output of this function will contain undesirable distortion due to the non-linearity. However, if the clipping is applied to a band-limited signal the harmonics occuring at multiples of the input components' frequencies can be made to lie outside the frequency range of the input. 12 CHAPTER 2. EARLY ALGORITHMS 13 output -T T Input Figure 2.1: Center-clipper transfer function In that case post-filtering with a filter whose passband is limited to the range of the input will remove the harmonic part of the distortion. It will not, however, affect any distortion components inside the input frequency range. 2.1.2 Implementation Erber and Gagne used the center-clipping approach in a two-channel system in which the stimulus was divided into two frequency bands: above and below 500 Hz. Each of the filtered signals was then subjected to center-clipping. In an effort to extend this approach I used a fourteen channel filterbank composed of third-octave five-pole Butterworth filters covering the range from 50 Hz to 4.5 kHz. Figure 2.3 shows the block diagram of the algorithm. Each of the bandpass signals was passed through a center-clipper with the threshold set to match the impairment. Subsequently each clipped signal was post-filtered to remove out-of-band harmonics. The pre- and post-filters were identical i.e. H,(w) = K(w). The procedure was implemented on a VAX 11/750 computer with the aid of SPUD, a signal-processing software package written at the Sensory Communications Group (13). As an initial test several sentences were processed and played out using a 16-bit DAC. The clipping thresholds were set to correspond to an auditory threshold of 60 dB SPL for all channels, with sentences presented at an overall input CHAPTER 2. EARLY ALGORITHMS 14 T -T r0\/ T ( A </' <I Figure 2.2: Effect of center clipping on sinewaves level of 80 dB SPL. 2.1.3 Results Upon listening to the output of the simulation it was discovered that while the processing certainly resulted in lowered intelligibility it also produced a noticeable and unexpected distortion that gave the output a "bubbly" quality. This effect is undesired since it is not believed to be a characteristic of actual hearing loss. The most probable cause of the "bubbly" artifact, we felt, was in-band intermodulation. Figures 2.4 and 2.5 illustrate this effect. The top panels show a test waveform and its spectrum, respectively. As is apparent from Figure 2.5 the waveform is composed of three sinewaves at 0.93, 1.0 and 1.04 kHz. The bottom panels show the result of center-clipping this signal CHAPTER 2. EARLY ALGORITHMS Filterbank CenterClippers -H (o .Input H 2((0) H 1((0) 15 - -+ T1 - T2 .. T1 Post-filters - -- ___ K, (w) +K2((0) output K 14((0) Figure 2.3: Center-clipping algorithm block diagram with clipping thresholds placed at 15, about half of the maximum amplitude. The clipped signal was post-filtered with a bandpass filter centered at 1 kHz with bandwith of 250 Hz, identical to the corresponding filter used in the simulation. As we can see from the spectrum of this signal the processing produced significant inter-modulation products which remain not only within the bandwidth of the filter but are intermixed with the original components. It is evident that no conventional filtering scheme will alleviate the problem. While this distortion would probably further degrade the intelligibility of speech, it is unlikely to correspond to actual features of an impairment. To ensure that the distortion was not an artifact of the sampled-data representation (i.e. aliasing), the test signals were upsampled by a factor of four and the simulation was repeated. There was no noticeable difference in the quality of the output. These observations lead to the conclusion that the undesirable bubbly CHAPTER 2. EARLY ALGORITHMS 16 Original Signal 30 15 0 a> 4-0 -15 E -30- Clipped and Filtered Signal 15 0 -15 0 5 10 15 20 25 Time, milliseconds Figure 2.4: Complex waveform illustrating the effects of clipping and post-filtering in one band effects, if they are due to intermodulation distortion, are an inherent effect of the processing. 2.2 Frequency domain clipping The preceding analysis suggests that the desired algorithm should suppress spectral components that fall below threshold without introducing new components. An obvious approach is to apply the center-clipping function of Figure 2.1 to the CHAPTER 2. EARLY ALGORITHMS 17 - 2000 1600 - Original Signal - 1200 800 - 0) CZ 400 -r-t I I 0 I ~J I I I I Clipped and Filtered Signal 400 200 iga -r 0 0 0.5 1.0 1.5 2.0 Frequency, kHz Figure 2.5: Spectra of waveforms from Figure 2.4 showing introduction of intermodulation distortion spectrum of the input rather than to the input itself. 2.2.1 Description Figure 2.6 shows the steps involved in this algorithm. To provide a running spectrum of the stimulus, Short Time Fourier Transforms (STFT) of length 25.6 ms (512 points for 20 kHz sampling rate) were computed. These were spaced 6.4 ms apart (i.e. they overlapped by 19.2 ms). Each STFT was subjected to center-clipping. CHAPTER 2. EARLY ALGORITHMS 18 Since the algorithm was implemented digitally this meant that each spectral component magnitude was passed through the function in Figure 2.1. The thresholds for different components could be set independently depending on the characteristics of the impairment. These modified STFTs (MSTFTs) served as input to the GriffinLim algorithm (5) whose output constituted the simulated signal. Input Short-Time -.- -- Fourier --.. o- T Transform STFTs Algorithm MSTFTs Griffin-Lim output T Spectral Clipping impairment thresholds Figure 2.6: Block diagram of the frequency domain clipping algorithm The Griffin-Lim algorithm iteratively reconstructs a signal from the magnitude of MSTFTs. It requires an initial estimate of the signal to be reconstructed. Since we want to preserve the same phase as in the original stimulus, the initial guess used was the unprocessed waveform. The procedure yields a signal such that the total squared error between the magnitude of its STFT and the input MSTFT magnitudes is minimized. 2.2.2 Perceptual effects The same sentences which were used to evaluate the time domain clipping algorithm were processed according to this algorithm. The spectral distortion was largely removed. However, the bubbly effect was only slightly diminished. In order to investigate the origin of this quality several test signals were processed. These were: 1) an amplitude-ramped square wave; 2) a frequency-modulated CHAPTER 2. EARLY ALGORITHMS 19 square wave; and 3) a frequency and amplitude-modulated square wave. They were chosen to test the theory that the effect which is mostly heard in the vowels is associated with the alteration of voice pitch. However, for none of these signals was the effect heard after processing. Further, examination of the spectrograms of the processed stimuli revealed nothing that would account for this effect. 2.2.3 Attempted improvements There was a possibility that the bubbly effect was due to time-aliasing. Clipping the STFT is equivalent to applying a time varying and oddly shaped but very steep filter. One would expect that in the time domain this implies a convolution with a long impulse response. It seemed plausible that a longer DFT than the one used originally would be necessary to accurately represent the desired output. To test this theory each 512-point frame was padded with zeros to 4096 points. Nonetheless, no significant changes in the output were observed. The Griffin-Lim algorithm never explicitly finds the inverse transform of the modified STFTs so the potential for time aliasing is probably smaller and less predictable. It appeared not to be a factor in this case. By a similar rationale the algorithm was altered to shift the transform windows by only 16 points. This would presumably improve resolution and more faithfully reproduce transitions in the speech waveform (the distortion was most apparent in the initial portions of vowels.) However, little change was observed in either the visual appearance or the sound of the output waveform. We don't have a good explanation for the bubbly effect. It might be partially related to the fact that the distance measure used by the Griffin-Lim algorithm to converge on the estimated output is a global one. Therefore the algorithm is prone to make "local" errors. For example, not infrequently the amplitude of the output of the simulation (after the spectral center-clipping) was observed to exceed that of the input in portions of the signal. Since both of the above algorithms showed this artifact no tests with subjects were carried out. The algorithm described in the next chapter turned out to merit closer inspection. Chapter 3 The AGC Algorithm This algorithm was based on the work of Vilichur described in Chapter 1. In a similar fashion it employs automatic-gain-control expansion to provide leveldependent channel gains. It explicitly simulates elevated thresholds, recuitment and reduced dynamic range. 3.1 Overview Figure 3.1 shows the block diagram of the algorithm. In the first stage the input is passed through a filterbank which divides it into N independent channels. Subsequent processing of all channels follows the same pattern. First, the short-time RMS level of the signal as a function of time is determined in each channel. The level waveform then serves as input to the gain computation block. Here, using parameters of the impairment, the relation of the input level to the impaired threshold is used to determine the appropriate attenuation for the particular channel. Since the level changes as a function of time the necessary attenuation will also change with time. Each channel signal is then multiplied by its computed gain signal. The modified channels are summed to produce the output. In the digital implementation used in this study the above processing applies to the input on a point-by-point basis. In fact, one may think of the gain computation block as creating a gain curve which then multiplies, point-by-point, the channel waveform. In the following section the implementation of the major functional blocks and the factors which influenced the choice of various parameters of the simulation are described. 20 CHAPTER 3. THE AGC ALGORITHM Level Detector Fiterbank ---a~~ 1 Gain Computation 2 channel 2 Level t wDetector --- HN (0 Gain Computation H()channel 1 Level Detector -- H2 (C) 21 Channel N otu Gain hComputation N _C Figure 3.1: Block diagram of the AGC algorithm 3.2 3.2.1 Functional blocks The Filterbank The bandwith of the filterbank was chosen to cover the frequency range from 50 Hz to 4.5 kHz. The upper cutoff was largely determined by the fact that available test stiun-Ai had been low-pass filtered at 4.5 kHz prior to digitization. The bandpass filters used were approximately third-octave except for the five at the lowest frequencies whose bandwiths were 100 Hz. With these requirements 14 filters were required to fill the desired span of frequencies. Table 3.1 lists the center frequencies and bandwiths of all filters. The filters were synthesized using a prototype three-pole Butterworth low-pass CHAPTER 3. THE AGC ALGORITHM Filter 22 1 2 Center Frequency (kHz) 0.100 0.200 Bandwidth (Hz) 100 100 3 0.300 100 4 0.400 100 5 6 7 8 0.500 0.625 0.800 1.025 100 150 200 250 9 10 11 1.300 1.600 2.000 300 300 500 12 13 2.550 3.225 600 750 14 4.000 1000 Table 3.1: Filterbank characteristics filter with cutoff at 25 Hz. Modulating its impulse response with cosine waves at increasing frequencies produced a series of filters with impulse responses hk(t) = 2h(t) cos(27rfAt + iAk), where hk(t) is the impulse response of the kth filter, h(t) is the impulse response of the prototype filter and fk is the desired center frequency of the resulting bandpass filter. Its bandwidth is twice that of the prototype, in this case 50 Hz. By adding the impulse responses of appropriate adjacent filters the filters of Table 3.1 were obtained. This synthesis method was chosen to take advantage of the approach developed by Schafer and Rabiner (15) which allows us to minimize the ripple in the combined frequency response of the filters in the filterbank. By appropriately choosing <kA, the phase of the modulating cosine, we may also be able to obtain a very nearly linear phase characteristic. In this case both of these goals were accomplished for <4k chosen such that the modulating cosines were all delayed by 1.3 ms; for example CHAPTER 3. (I = -58.5* THE AGC ALGORITHM 23 where f2 = 125 Hz. 2 0 M V - - CZ - -4 -------- --- -6 -8 -101' 0 1.0 2.0 3.0 4.0 5.0 Frequency, kHz Figure 3.2: Total magnitude of the filterbank's frequency response Figure 3.2 shows the magnitude of the frequency response of this filterbank. It is apparent that the maximum ripple is much reduced from the simple filterbank used initially (see Figure 4.5), from about 5 dB peak-to-peak to only about 0.5 dB. Figure 3.3 illustrates how successful the filterbank was in achieving linear phase. The top panel shows the waveform of the syllable /ba/. The bottom panel shows the result of passing this waveform through the filterbank, summing the outputs and shifting by 1.3 ms to compensate for the expected delay of the filters. The filterbank clearly preserves the fine structure of the signal. Although the phase of a sound stimulus is widely regarded as relatively unimportant, there is no guarantee that it is also unimportant when the signal is subjected to nonlinear processing. Through careful design of the filterbank we are assured that the simulation will be free of artifacts from at least that part of the processing. CHAPTER 3. THE AGC ALGORITHM 24 Original 100 60200 -20-603 -100- Filtered 100 60200-20-60-100- I 0 0.2 I 0.4 I 0.6 I 0.8 I 1.0 Time, seconds Figure 3.3: Effect of the filterbank on the syllable /ba/ 3.2.2 Level detector The level detector's function is to provide a running estimate of the intensity in the input signal. For the purposes of this algorithm it is important that the level detector track variations within syllables (tens of milliseconds) but not track variations at the fundamental frequency rate (10 msec or less). To implement this function, the instantaneous power Ek[n] in a given channel k is first found by simply squaring the channel signal ck[n], i.e. Ek[n] = Ick[n]1 2 . These powers then have to CHAPTER 3. THE AGC ALGORITHM 25 be smoothed to produce the short-term band levels ek[n]: ek[n]=10log z w[n-m]E[m] mn=-oo The function w[n] is the smoothing window. In this case this was chosen to be a noncausal rectangular window of length 20 msec (effective bandwidth ~ 50 Hz). As reported by Bustamante (1) this window successfully eliminates pitch related fluctuations while preserving syllabic variations. By using a window which is symmetrical in time we also eliminate the delay between the level-detector output and the input signal. 3.2.3 Gain Computation This is, arguably, the most important part of the algorithm. By appropriately choosing the transfer function of this block we seek to reproduce the elevated thresholds and abnormally steep loudness growth (recruitment) observed in hearing impairments. Description Figure 3.4 illustrates the approach. The dashed curve shows loudness growth experienced by normal listeners as a function of the stimulus intensity. In general this will be a straight line above the normal threshold of audibility labelled T. Typical impaired loudness growth as might be measured with loudness balancing is represented by the heavy line. At impaired threshold (Ti) the perceived loudness equals that of the normal listener at his threshold. For the range of stimulus intensities between T; and T, the loudness grows faster than linearly. This is the phenomenon of recruitment, therefore T,. will be called the threshold of recruitment. For intensities above T, impaired loudness growth matches the normal curve. This implies so-called complete recruitment. In order to duplicate for the normal listener the impaired loudness growth the algorithm has to scale the input signal such that the output intensity is related to the input by the bold line. The double arrow indicates the amount of attenuation CHAPTER 3. THE AGC ALGORITHM A 26 output intensity (dB) Tr E -0 ~- 7 - - - attenuation factor +n i Tr input intensity (dB) Figure 3.4: Gain computation transfer function. See text for details (in dB, at a particular input intensity) necessary to accomplish this. As expected, the attenuation decreases with increasing input level until at T,. the gain becomes one. The attenuation used for signals whose level falls below threshold has to bring them below the normal threshold. The exact value is unimportant. In practice the algorithm essentially throws away most of the signal except for the 5 dB range immediately below threshold. In this range the attenuation increases significantly but some signal is still passed through. This is done to avoid a possible abrupt onset of the signal when its level crosses the threshold which might be audible as high frequency transient distortion. This function may be easily adapted to include lowered threshold of discomfort experienced by some hearing impaired listeners (17). All that is necessary is to CHAPTER 3. THE AGC ALGORITHM 27 extend the curves further. At some point below the normal threshold of discomfort the impaired curve will begin to exceed the normal curve. Now the distance between the curves may be taken as the amount of amplification which will bring the stimulus up to the normal discomfort threshold. Due to difficulties with a consistent definition of discomfort threshold this aspect of the simulation was not tested. Given this method, the algorithm needs two impairment parameters for each channel: threshold of audibility and threshold of recruitment. Throughout this study tone detection thresholds were used for the former. The choice of the latter is described in the following section. Characterization of recruitment Carver (2) describes four patterns of recruitment that have been observed. Of these, the linear pattern of Figure 3.4 and the asymptotic pattern are by far the most common. They tend to differ realtively little and only at the end of the recruitment range. Stevens (18) suggests that there is only one pattern and the differences in shape are a result of averaging. Only linear recruitment was simulated here. Ideally, the threshold of recruitment would be obtained from actual measurements of the impaired function. and not measured. In practice, these data are often not available More importantly, for bilateral impairment with thresholds above normal for all frequencies it is impossible to use the balance test to measure the recruitment. Other methods are being developed (7) but none is yet widely available. Hallpike and Hood (6) have measured recruitment in over 4000 subjects. Their data strongly suggest that the range of recruitment is a fairly orderly function of the degree of loss. Their observations, which are in general agreement with those of Miskolczy-Fodor (12), are summarized in terms of the angle (steepness) of the recruitment curve. This angle a (in degrees) between the recruitment function and horizontal axis is well approximated by a = 47 + 0.45(T 1 - Tn) where Ti and Tn are in dB, which predicts increasing slope for more severe im- CHAPTER 3. THE AGC ALGORITHM 28 pairments, as commonly observed. After appropriately converting this metric to range of recruitment (by simple geometric relations) we obtain the threshold of recruitment as follows: T = tan(a)(T - Tn) tan(a) - 1 For example, for Tn=10 dB SPL and T=70 dB SPL, T,.=94 dB SPL (recruitment range of 24 dB). Figure 3.5 shows the recruitment range as given by the above equation as a function of the impaired threshold. The normal threshold was assumed to be 0 dB SPL for this plot. The plot will in fact remain the same if the abscissa is taken as the difference between the impaired and normal thresholds. I I I I I I I I I 40 -o 30 1 Co C 20 Ei =3. 10 01 0 20 40 60 80 100 Impaired Threshold, dB SPL Figure 3.5: Recruitment range as a function of T; T,., assumed to be zero For very small and very large impairments unrealistically small recruitment ranges are predicted. In order to avoid this problem the algorithm allows a minimum recruitment range of 15 dB. CHAPTER 3. THE AGC ALGORIT HM 29 This method of calculating the recruitment range was chosen since it is believed to be a more accurate representation of the loudness growth than the criterion employed by Villchur as described in section 1.2.4. In that study (20; 21) a single expansion ratio is computed for each frequency band based on the ratio of the normal span of hearing to the impaired span of hearing. These are defined in each band as the difference between the threshold of audibility and an equal loudness contour. For normal subjects this contour is chosen as the 74-phon equal loudness contour. The impaired contours shown by Vilichur are generally about 20-30 dB above that. For impaired and normal thresholds of T and T, respectively and the equal loudness levels of Xi and X, the expansion ratio is (X, - T.)/(X - Ti). After the input signal's level is expanded with this ratio, attenuation is used to simulate the elevated threshold. Figure 3.6 shows an example of the resulting gain curve (bold line) based on thresholds at 1000 Hz in one set of charcteristics shown by Villchur. Here T = 55 dB, X = 90 dB, T, = 8 dB and X, = 74 dB. It is contrasted with the gain curve which would be employed by the AGC algorithm under these conditions. The AGC curve shows steeper loudness growth as well as saturation, i.e. complete recruitment and normal loudness growth above a certain threshold. Villchur's expansion is shallower and continues indefinitely resulting in overrecruitment. These differences are typical. Based on available data we believe the AGC algorithm reproduces abnormal loudness growth more faithfully. 3.3 Implementation The simulation was implemented on a VAX 11/750 computer. C-language program which includes all of the functions of Figure The code of the 3.1 is included as Appendix A. The filterbank was synthesized beforehand using SPUD (13) and the impulse responses were stored for access by the program. Filtering was done by means of Fast Fourier Transform. The input stimuli were either digitized and stored before processing or generated digitally. Likewise, the output of the simulation was in digital form. The signals were played out using a 16 bit Digital-to-Analog converter and a Crown adjustable amplifier which set the output level and acted as CHAPTER 3. THE AGC ALGORITHM 30 output level (dB) - - 100 80 - 120 - -I 20 - 40 -' -' - 60 I 20 40 60 80 I I W, 100 120 input level (dB) Figure 3.6: Comparison of the AGC algorithm gain curve and that of Villchur's simulation the headphone driver. For all tests the output was presented to TDH-39 headphones mounted in Grason-Stadler .001 circumaural cushions. The full implementation is diagramed in Figure 3.7. The DAC's input range was -32768 to +32767. To minimize quantization noise the stimuli were kept as close to full range as practical and the overall level was adjusted externally. The DAC output 0.1526 1V/ unit and the phones nominally produced 6.325 mPa / mV. Therefore the relation between the digital signal level DA, analog gain G and the sound pressure level at the headphones, SPL,, was: SPLout = Dk + G + 33.7 where all quantities are in dB. The simulation program takes this into account by CHAPTER 3. THE AGC ALGORITHM 31 VAX sound A/D Disk -WSimulation Program :PDAC Gain TDH-39 Figure 3.7: Implementation block diagram taking the intended sound level of the input as a parameter. It then calculates the necessary analog gain and converts the impairment thresholds from dB SPL to the digital signal levels. For example: a digitized syllable with RMS overall level of 3276 (a tenth of full range) has D1 = 70.3 dB. If it is to be played out at 72 dB SPL then G = -32 dB. A threshold of 60 dB SPL will then map to a "digital" threshold of 58.3 dB. Naturally both the unprocessed and processed stimuli have to be played using the same analog gain setting (unless digital scaling is used as well). 3.4 Illustration Figures 3.8 and 3.9 show an example of the effect of the algorithm on the syllable /bi/. The top panel of Figure 3.8 shows the unprocessed digitized utterance. Its RMS level is 1638. This syllable was processed by the algorithm assuming an overall output level of 94 dB SPL and using the thresholds of subject PG (see Figure 5.1.) The output of the simulation is shown in the lower panel. Figure 3.9 illustrates the steps in the processing. The top panel shows one channel of the input (i.e. the syllable), in this case channel 3: frequency range 250 to 350 Hz. The second panel shows the output of the level detector. Note that the scale has been converted to dB SPL given the assumptions of the preceeding paragraph. The impaired threshold in this frequency band is 76 dB SPL, the normal threshold is 23 dB SPL. This implies that the threshold of recruitment is 104 dB CHAPTER 3. THE AGC ALGORITHM 32 Unprocessed /bi/ - 6k I I I - 3k I L [[[III I Ia 0 + -o II1.1I 1. III II E -3k I I I - I II 'F" I I I'll liii.' 1k I, 0 Simulated /bi/ 111.6 L vp -1k 0 100 200 300 400 Time, milliseconds Figure 3.8: Effect of the AGC algorithm on the syllable /bi/. See text for relevant parameters CHAPTER 3. THE AGC ALGORITHM 33 SPL. Since the peak level is only about 99.5 dB normal loudness is never reached with this simulation in this channel. The third panel shows the output of the gain computation block for this channel. As expected, the gain is zero for those parts of the input where the level falls below 76 dB. The bottom panel shows the output for this channel obtained by multiplying the channel input by the gain curve. CHAPTER 3. THE AGC ALGORITHM 34 input - 3k - 1k 0- -1 k -3k -J a(/, m 2 80 level 40 V 0 00.3 Cu - gain - output 0.2 0.1 0 1k E 0 -1k 0* 100 200 300 400 Time, milliseconds Figure 3.9: Steps in the simulation of the syllable /bi/: channel 3, T= 76 dB SPL, T, = 23 dB SPL Chapter 4 Pilot Evaluation of AGC Algorithm 4.1 Aim As an initial test, the AGC algorithm was used to simulate the impairments described by Zurek and Delhorne (24). They compared consonant reception in noise by hearing impaired listeners to that of normal subjects whose thresholds were elevated by the addition of spectrally shaped noise. In that study subjects were placed in one of five categories according to the shape of their audiograms. For each category of impaired subjects an "average" audiogram was determined. The normal subjects' thresholds were then made to match the impaireds' by varying the spectral shape of the masking noise. The losses were mild to moderate, all less than 70 dB. The intensity of masking noise necessary to simulate more severe impairments would have been uncomfortable. In the current study the same stimuli as those used by Zurek and Delhorne were altered according to the AGC algorithm to simulate each of the five average loss patterns. The performance of two normal hearing subjects with these simulation stimuli was compared to that of the original hearing impaired subjects as an initial indicator of the accuracy of the simulation. 4.2 4.2.1 Methods Speech materials The stimuli used were the same consonant-vowel syllables (CVs) used by Zurek and Delhorne. Twenty four consonants (b,tf,d,f,g,h,d3 ,k,l,m,np,r,s,J, t,0,v,w,hw,5,y,z,3) were paired with three vowels (a,i,u) for a total of 72 syllables. They were spoken by 35 CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM 36 a male speaker and recorded in a quiet/anechoic setting. These recorded syllables were then low-pass filtered at 4.5 kHz, sampled at 10 kHz, and stored on a computer disk. They were played back using a 12-bit D/A converter and a 4.5 kHz low-pass filter. Figure 4.1 shows the block diagram of the experimental setup used in the previous study for testing the hearing impaired subjects and noise-masked normals. The syllables were presented to the point labeled "CVs" in that diagram. Figure 4.2 shows the block diagram of the processing used in the present experiment; here the syllables were introduced at the "CVs" node still in their digitized form. 0 CVs Frequency-Gain Characteristic SNR Overall +Multifilter G a In Level Gain SSN Masked Normal -White Noise Multifliter Threshold Noise Shaping r TDH-39 paired Figure 4.1: Block diagram of Zurek and Delhorne experimental setup CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM Frequency-Gain Characteristic Cvs S NJRSNR+Filter Digital SSGain Simulation Program 37 Overall Level Digital Gain DAC Gain TDH-39 Figure 4.2: Block diagram of the pilot experiment setup 4.2.2 Speech-spectrum noise and frequency-gain characteristics In the Zurek and Delhorne study the interference noise was generated by analog equipment and added to the CVs during presentation. This was continuous Gaussian noise, filtered to have the spectral shape of babble, hence the designation "speech-spectrum noise" (SSN). For the present implementation of this simulation the noise had to be added to the CVs while they were in digital form. White Gaussian noise was generated digitally and spectrally shaped with a filter whose frequncy response magnitude, shown in Figure 4.3, was chosen to match the spectrum of the SSN used by Zurek and Delhorne. Two frequency-gain characteristics were employed. For the flat characteristic uniform gain was applied to the CV-plus-noise stimulus. For the high-frequency emphasis (HFE) characteristic the gains of the multifilter used by Zurek and Delhorne were chosen so that the SSN was transformed into pink noise. For the present study the syllables were digitally filtered prior to simulation with a filter whose frequency response magnitude, shown in Figure 4.4, was chosen to match Zurek and Delhorne's HFE filter. CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM 40 I I I I I I I I I I I I 38 I 30 20 V (D V: 0M 10 0I -10 -20 -30 0 I 1.0 I 2.0 I 3.0 4.0 I 5.0 Frequency, kHz Figure 4.3: Frequency response of speech-spectrum noise shaping filter 4.2.3 Simulation parameters Since this evaluation was performed relatively early in the development process several parameters of the algorithm used in the simulation were different from those characterizing the final version as outlined in the previous chapter. The smoothing window (for level computation) was 10 ms long as opposed to 20 ms eventually decided on. The filterbank used was simply a collection of third-octave fifth-order Butterworth filters (slopes = 30 dB/octave) with response magnitudes intersecting at half-power points. The frequency response magnitude of this filterbank is shown in Figure 4.5. It has larger ripple than the filterbank used in the final version. Finally, the recruitment range was a uniform 30 dB for all channels regardless of the magnitude of the loss. The shorter smoothing window is unlikely to have had a large effect. It was observed that the shape of the waveform, particularly the consonant part, was not greatly affected by changing the window to 20 milliseconds. Potentially more CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM 39 16 1 21 -0 8 a; -' CT) :|2 4 0 -4 -8 C 1.0 2.0 3.0 4.0 5.0 Frequency, kHz Figure 4.4: Frequency response of the high-frequency emphasis filter significant problems arise from the ripple since it produces varying amplification near the overlap frequencies. The recruitment range, while admittedly somewhat arbitrary, is consistent with the average ranges observed (18; 19). Furthermore, the thresholds simulated are averages of several audiograms and so it is not unreasonable to use an "average" recruitment range. 4.2.4 Subjects The impaired subjects are described in detail in the aforementioned paper. Figure 4.6 shows average tone-detection thresholds of the subjects in each category. These average thresholds were used in the simulation. For this pilot study the author and another member of the laboratory served as the normal-hearing subjects. Tests in the laboratory confirmed that their thresholds were within the normal range. CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM 40 5 0 M -5 c0 C 10 15 -20 0 1.0 2.0 3.0 4.0 5.0 Frequency, kHz Figure 4.5: Total frequency response of the filterbank 4.2.5 Procedure The conduct of the experiment is described in Zurek and Delhorne. It is repeated here for convenience and because it was used again (with minor modifications) in the experiments discussed in the next chapter. Unless otherwise noted the following applies to both the previous and current studies. The subjects were tested while seated before a computer terminal inside a soundproof room. Each experimental run comprised the presentation of a set of 72 CVs at a particular SNR, frequency-gain characteristic, and overall level. The subject's task was to identify each syllable as it was presented. time limit for responding. There was effectively no Acceptable responses were in the form of an English alphabet code for the consonants and the characters "a", "i", and "u" for the vowels. The code followed the IPA alphabet except where the latter contained nonEnglish characters. For those consonants the codes used were: ch=/tf/, j=/d3/, sh=/f/, th=/0/, wh=/hw/, xh=/6/, and zh=/ 3 /. To avoid possible confusion the CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM I - 100 I I I 41 1 GrouplI, 3Ss Average Threshold 50 - 4%4.. -. 0100 Normal Threshold 00 - -J C0 . Group 111, 3 Ss r, .% .% 50- 50 GroupllV, 2Ss 0 0- -F- 100 C 00 - Grou p V, 5 Ss 50 50 - (I) 4% 0 0I I I 0.25 1.0 4.0 I I 0.25 1.0 Frequency, kHz Figure 4.6: Tone detection thresholds of the impaired subjects 4.0 -.C CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM code for /j/ 42 was "y." The subjects had available during the experiment a list of all valid responses along with illustrative words containing the sounds. Both normal subjects were familiar with the response code. Zurek and Delhorne reported that their subjects seemed to have no problems mastering it. No trial-by-trial feedback was given. The next syllable was presented 0.5 seconds after the response to the previous stimulus. When invalid responses were made the subjects was prompted to choose an acceptable answer (i.e. one in the response set). The subject could waive responding on any trial by typing "return." The stimulus presented on that trial was returned to the pool of unplayed syllables to be replayed later in the same run. In this way valid responses for all syllables were obtained. 4.3 Results Figures 4.7- 4.11 show the percent of consonants identified correctly as a function of SNR for each of the five impairment categories. The SNRs employed in the present study were -10, -5, 0, and 5 dB as well the "Quiet" (Q) condition: SNR > 20dB. The bold solid lines and the bold crosses represent the average performance of the two normal subjects. For the flat frequency-gain characteristic one of the two normal subjects (PD) was tested once at each SNR and overall level while the other (PZ) was tested once in "Quiet" and at -10 dB SNR. For the HFE characteristic both normals were presented stimuli once in the Q condition and once at 0 dB SNR for each overall level. As can be observed from these graphs, the current simulation produces identification scores which, in most cases, closely match those achieved by the impaired subjects. It can be safely said that it is at least as accurate in predicting the performance of the impaired subjects as the masking noise employed in the previous study. The results were judged to be promising enough to warrant the more rigorous and extensive investigation that is described in the next chapter. It also appears from these experiments that the present simulation gives at least as good or somewhat better performance than noise masking. This may indicate that noise masking has other effects detrimental to intelligibility in addition to CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM elevating thresholds and producing recruitment. 43 CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM I I I I I - 100 50 - Flat, 72 dB SPL U) - 80 C C 0 0 130 Impaired C/ C 0 0- ------ Noise Simulation 1000 HFE, 96dB SPL C-) Of Current Simulation 110 02 50JAA 0 'g 0- I - -20 I 0 20 Q SNR, dB Figure 4.7: Consonant identification scores for Group I 44 CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM S. I F I 45 I - 100 Flat, 72 dB SPL A Impaired 03 ------ Noise Simulation 50C,) - Current Simulation 4. C 0 -I 0100- I I I -100 Flat, 84 dB SPL. . I ID C 0 HFE, 96 dB SPL El +I ,' '',t -' - 50 ' 0 - 50 B 03 0- -20 I ,I 0 , ,2 20 I1- Q -20 , I 0 SNR, dB Figure 4.8: Consonant identification scores for Group II K0 20 Q CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM L I I I ,I i 100- Flat, 72 dB SPL 13 0 0 El - 50 C/) itL - 0 - 0A Impaired 0 C/) C 0 - 4- 100 ------ Noise Simulation - 00 HFE, 86 dBL C) t __ o 0 AX / - 50 Current Simulation ' 0 B3 0-20 0 20 Q SNR, dB Figure 4.9: Consonant identification scores for Group III 46 CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM I I III I i - 100 Flat, 72 dB SPL - 50 0 C, a C: ,-,---..---- 4-' 0 Cd C,) 0 Impaired 0 *O ------ Noise Simulation 0100- Current Simulation HFE, 96 dB SPL 0 U' 01 - - 0. E - 50 I'- - = -= = 00 0 ,'6'' 0 -20 0 I 20 i Q SNR, dB Figure 4.10: Consonant identification scores for Group IV 47 CHAPTER 4. PILOT EVALUATION OF AGC ALGORITHM - I I I I I 48 I. 100- A v o Flat, 84 dB SPL _ - - Impaired 3 50- - ------ Noise Simulation C (Ti C: Current Simulation o - # Cn 100- -100 Flat, 94 dB SPL HFE, 96 dB SPL 0 01 o --,---- 0 -20 -- #-- -0 0 L0 0 20 Q-20 0 20 SNR, dB Figure 4.11: Consonant identification scores for Group V Q Chapter 5 Detailed Evaluation of AGC Algorithm 5.1 Aim This series of experiments was conducted to provide a rigorous comparison of the speech intelligibility performance of hearing impaired subjects with that of simulated normals. The goal was to find to what extent the AGC algorithm duplicates reduced intelligibility and to gain insight into any deficiencies it might have. The experiments included a consonant identification test similar to that described in the previous chapter as well as sentence comprehension. The stimuli were presented to three hearing impaired listeners whose losses may be classified as severe. The magnitudes of these losses were out of the range that can be simulated with masking noise. Their performance was compared to that of three normal hearing subjects listening to the same stimuli passed through the AGC simulation. Some of the CV tests were repeated using the simulation without including recruitment. This was done in order to determine the relative importance of recruitment in determining intelligibility. 5.2 5.2.1 Methods Speech materials For the consonant identification experiments the same set of consonant-vowel syllables as used in the pilot study was employed. Their presentation followed exactly the procedures described in section 4.2. The sentences were taken from the Harvard-IEEE sentence lists (9). They were spoken by a male and recorded in the same anechoic setting as the CVs. 49 CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 50 The recorded sentences were low-pass filtered at 4.5 kHz and sampled at 10 kHz. Their processing and playback followed the same pattern as the CVs (substitute "Sentences" for "CVs" in Figure 4.2.) Speech spectrum noise was digitally added to the stimuli as before. Two settings of SNR, the "Quiet" and 0 dB were used in combination with the same two frequency-gain characteristics (flat and HFE) to yield four stimulus conditions, in addition to overall level. 5.2.2 Simulation parameters For the complete simulation all parameters were set to the values given in Chapter 3. In particular, rather than using a fixed value, the thresholds of recruitment were computed for each impairment and frequency band. The discomfort thresholds were not simulated i.e., the simulation provided unity gain for all signal levels above the recruitment range. The simulation was altered for the final set of experiments where no recruitment was simulated. This simply involved changing the transfer characteristic of the gain computation block in Figure 3.1. The curve above threshold was changed to have a slope of one. This implied simple and constant attenuation of the signal by the difference between the impaired threshold and the normal threshold. Section A.2 lists the minor alterations to the code necessary to effect this function. 5.2.3 Subjects Impaired The three impaired subjects had previously participated in experiments in the laboratory and their names were obtained from group files. One of them (PG) had extensive prior experimental experience. All normally wore a monoaural hearing aid in the right ear, which was also the ear tested; the aids were not used during testing. Audiometric records (less than one year old) showed those hearing impairments to be bilaterally symmetrical and recruiting. Other relevant chracteristics are summarized in Table 5.1. CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM Subject PG AL Age 49 23 Sex M F PTAt 66 71 Comments Longstanding Longstanding PB 37 F 76 Congenital t 51 The average hearing loss at 500, 1000 and 2000 Hz in the test ear Table 5.1: Impaired subject characteristics Threshold testing was performed in the laboratory. An adaptive forced-choice psychophysical precedure (10), which targets the 71% correct signal level, was used to measure thresholds for 500-ms pure tones at octaves from 125 to 4000 Hz (standard audiometric frequencies). The threshold measurement was performed twice. The average of the two measurements was subsequently used to determine thresholds in the 14 frequency bands used by the simulation. These thresholds were determined by linear extrapolation from the measured thresholds. Figure 5.1 shows the thresholds of the impaired subjects. Normal Three normal subjects who had previously participated in psychophysical experiments in the laboratory were recruited. All three were female, ages: KG-44, AW38, DW-41. Their thresholds were measured using the same procedure as for the impaired subjects. They were found to lie within 15 dB of normal for all frequencies tested. None of the subjects (normal or impaired) had listened to Harvard sentences before this study. 5.2.4 Procedure CV Syllable tests The details of the conduct of this test were essentially unchanged from that described in section 4.2.5. The syllable sets were presented in groups of four (four CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM PB -J 100 0L I I I Impaired Thresholds - I_ -~ AL Cl) -PG ~0 .c0 (D II 52 -- 50 FNormal -- 0 .25 1.0 4.0 Frequency, kHz Figure 5.1: Impaired subject thresholds conditions for a given overall level). The presentation of successive sets was automatic and the subjects were free to take breaks between sets. For the simulated normals consecutive sets always simulated the same impairment. In practice that meant one impairment was simulated per session. A typical session lasted 2 to 3 hours. Sentence tests The sentences were presented in groups of 10. They were presented in a setting much like the syllables. The subjects were prompted on the terminal screen to type "return" to hear the next sentence. After hearing a sentence they were to write it down on a pad of paper using underscores for words heard but not understood. After the first presentation of a sentence and the subject's response the same sentence was played once more and the subject was asked to identify it again. This was done to CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 53 compensate for the probable advantage of the hearing impaired subjects who would be more accustomed to the reduced intelligibility of the sentences. Therefore a run comprised 20 utterances. No feedback was given during or after a run. The scoring scheme is described in the results section. 5.3 5.3.1 Results CV Syllables The performance of the impaired subjects was compared to that of the simulated normals both in terms of the average percent of consonants identified correctly and in terms of the pattern of errors, as summarized by an analysis of transmitted features. Percent consonants correct For each subject and each listening condition the experiment was performed at least seven times (using exactly the same speech tokens); the first two runs were treated as training and their results were not included in the calculations. It was observed that the performance of the subjects reached a fairly constant level after the practice tests. Consequently, each datum represents the average of at least five experimental outcomes. Figures 5.2 through 5.6 present the percentages of consonants identified correctly by the impaired subjects and by the simulated normals for particular stimulus levels. These levels were selected individually for the three impaired listeners to achieve scores well above zero but below 100% correct. The x-axis lists the listening condition i.e. combinations of frequency-gain characteristic and signal-tonoise ratio. The dark bar represents the impaired subject while the cross-hatched bars show the scores of the normals. The error bars indicate the standard deviation of the scores. CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 54 80 60 0 0 0 40 El -xi C 0 0. IN PG DW KG AW 20 0 flat, Q flat, OdB HFE, Q HFE, OdB Condition Figure 5.2: Consonant identification scores for impairment PG; 94 dB SPL overall level 80 60 0 AL C 0 DW E E KG E AW 40 20 0 flat, 0 flat, OdB HFE, Q HFE, OdB Condition Figure 5.3: Consonant identification scores for impairment AL; 94 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 55 80 - 60 HAL 40 -DW 0KG C ~AW It 20- 0 flat, 0 flat, OdB HFE, Q HFE, OdB Condition Figure 5.4: Consonant identification scores for impairment AL; 100 dB SPL overall level 80 60PB 40- iDW C I. 0KG EAW 20 0 flat, Q flat, OdB HFE, Q HFE, OdB Condition Figure 5.5: Consonant identification scores for impairment PB; 98 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 56 80 60 0W. H FE, 40, H PB ~DW 0. 20 0 flat, Q flat, 0dB HEE, Q HFE, 0dB Condition Figure 5.6: Consonant identification scores for impairment PB; 104 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 57 The data show the following trends. First, for the flat frequency-gain characteristic the matching of scores between the impaireds and the simulated normals is very close. The notable exception to this pattern is impairment PB at 104 dB SPL (Figure 5.6) where simulated normals' scores exceed the impaired's score by 5-15 percentage points. Second, the normals perform better under HFE conditions. The exception to this pattern is PG at 94 dB SPL, where the match is very close. Further, the discrepancy with HFE increases with increased severity of the impairment (from PG to AL to PB) and with the level of the stimulus. For subject AL, under the HFE conditions the normals perform better by an average of 8 percent at an overall level of 94 dB and 14 percent at 100 dB. For subject PB, the differences were 20 percent at 98 dB and 25 percent at 104 dB. Feature transmission Appendix B lists the consonant confusion matrices obtained from the above experiments. Direct comparison of error patterns in such matrices is difficult to perform particularly with the response bias which was present in some cases. Moreover, there was an insufficient number of responses per cell to provide reliable estimates on which to base simulation analyses. Thus we opted for the SINFA analysis (23) used by many previous investigators. Table 5.2 lists the features used in the analysis as they apply to the consonants included in the stimuli. In the analysis the amount of information available in each feature for the given stimulus set is computed. From the confusion matrix the percent of information in each feature transmitted by the subject is computed. This was done for each subject for every condition (SNR, gain-characteristic, overall level) tested. The matrices from individual runs were pooled together prior to this procedure yielding matrices with each consonant presented at least 15 times. If the algorithm simulated hearing impairment perfectly, the percent of information transmitted for each feature would be the same for the impaired subject and the simulated normals. In order to assess the correspondence of feature transmissions a correlation coefficient, was computed. This coefficient, Cf, was defined as follows: CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 1 +- + + B D G + 1 + + 1 + + 1 +r fricative TH S SH 2 2 2 2 + +- + + 2 3 4 V XH - + Z + ZH CH J M N R + + + + 3 3 4 4 5 5 5 5 5 + 2 2 2 2 + + + + 2 3 4 5 + 1 4 4 1 4 4 5 ling-detal Y H HW 5 2 2 + 1 4 M M + + IN R. 5 6 + + + ling-palatal + P M + B P -+ M P B + + + + + ling.velar + 5 1 + ling-alveol + 4 1 + labi-dental M M + + B P + M M + + B B B P P M B B Y H B + -+ back DURATION 4- + + 4- 41P T K B D G F TNT S SH V XH + -+ Z ZN + C11 + + + - middle + + PLACE?3 front + + bilabial L + nasal sonorant PLACE-6 W + affricate P + K 1 + T 1 + stop JP + VOICING MANNER + Peature 58 J M W L H Table 5.2: Feature table for Twenty-Four-Consonant set Cf = EM=J~i[M1- Z fA)(fn[m] (f[m] fi - j;) 2 Z - in) I(f [m] - f)2 where M is the total number of features, f 1 [m] and fn[m] are the mth feature transmission percentages for the impaired subject and simulated normal subject, respectively, and A, and In are the average feature transmission percentages for the impaired and normal subjects. A C1 of 1 indicates complete correlation, i.e. the transmission percentages of the impaired and normal subject are multiples of each other by a constant ratio. If plotted versus one another they would lie on a straight line. A coefficient of zero indicates that the data are uncorrelated. Figure 5.7 gives an example of a condition for which the correlation between impaired and normal listeners was 0.84 The symbols plot the percent of information transmitted by the impaired subject PG for each of the 19 features versus the analogous score for the simulated normal KG. The listening condition was in quiet, with flat amplification and overall level of 94 dB SPL. The solid line shows unity slope. This plot shows a qualitative agreement between feature transmissions typical of correlations in the 0.80 to 0.90 range. CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 59 100 Flat Quiet, 94 dB SPL 80 EC,, C: CI 60I C 0 00 40I 0 C 0 00 0 20 (5 0 - 00 0 20 0 40 60 80 100 PG % Information Transmitted Figure 5.7: Example of feature transmission matching; see text for details Tables 5.3- 5.7 list the correlation coefficients between each impaired subject and the three simulated normals at all listening conditions. The feature transmission correlations show generally very good agreement for all subjects and conditions simulated. The lowest correlation coefficients occur at one stimulus level for the HFE condition with impaired subject PG (Table 5.3). This is somewhat surprising since in this case the overall scores of the normals were closely matched to those of the impaired listener. 5.3.2 Sentences The performance of the subjects on the sentence comprehension task was evaluated using a dual approach as well. First, the raw score: percent of keywords identified CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE, 0 dB SNR Correlation DW KG 0.80 0.84 0.85 0.89 0.77 0.55 0.78 0.65 60 (C,) AW 0.89 0.89 0.50 0.58 Table 5.3: Feature transmission correlation coefficients; impaired subject PG, 94 dB SPL overall level Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE, 0 dB SNR Correlation DW KG 0.85 0.94 0.92 0.85 0.90 0.89 0.96 0.89 (C,) AW 0.89 0.82 0.89 0.89 Table 5.4: Feature transmission correlation coefficients; impaired subject AL, 94 dB SPL overall level Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE, 0 dB SNR Correlation DW KG 0.82 0.87 0.87 0.91 0.94 0.91 0.91 0.90 (C,) AW 0.84 0.86 0.92 0.85 Table 5.5: Feature transmission correlation coefficients; impaired subject AL, 100 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM Correlation DW KG 0.90 0.89 0.68 0.85 0.81 0.89 0.80 0.80 (C1 AW 0.87 0.77 0.79 0.77 ) Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE, 0 dB SNR 61 Table 5.6: Feature transmission correlation coefficients; impaired subject PB, 98 dB SPL overall level Correlation (C1 DW KG 0.91 0.91 0.87 0.90 0.76 0.73 0.83 0.87 ) Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE, 0 dB SNR Table 5.7: Feature transmission correlation coefficients; impaired subject PB, 104 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 62 Impaired # Subjects PG AL per Condition 30 40 flat, quiet 89 94 flat, OdB SNR 94 98 HFE, quiet 89 94 HFE, OdB SNR 89 94 PB 30 104 104 104 104 of Sentences Overall Level (dB SPL) Table 5.8: Sentence characteristics by subject and listening condition correctly, was determined. Second, an attempt was made to determine the degree to which inter-subject errors were correlated. Percent Correct Table 5.8 lists the levels at which the sentences were presented to the impaired subjects and the number of sentences presented for each listening condition. The stimuli were then passed through the simulation and presented to the normal subjects. The normals, therefore, heard the same sentences (albeit processed) as the impaireds. Different sentences were used for each of the three impaired subjects. Each of the sentences used in this study contained fie keywords. The subjects' responses were scored for the number of these keywords they identified correctly. Mistakes in tense or number were not counted as errors. The sentences were presented under the same four listening conditions as the CVs. An attempt was made to set the overall level of the sentences so that the impaired subjects' scores were about 50% correct for maximum sensitivity. Only the second response for a given sentence was used. Figures 5.8- 5.10 show the overall scores obtained by the impaireds and the simulated normals. Each score represents the percentage correct out of 30 sentences containing 150 keywords for impairments PG and PB and 40 sentences containing 200 keywords for impairment AL. CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 1nn I 800 C.) 60- U (I) 0 40- PG DW KG AW g 20U flat, Q flat, 0 dB HFE, Q HFE, 0 dB Condition Figure 5.8: Keyword identification for impairment PG 100 *1 20- ["Iii 0 --- 80 60 0 40 - NO OR flat, Q flat, 0 dB HFE, Q EU 0 HFE, 0 dB Condition Figure 5.9: Keyword identification for impairment AL AL DW KG AW 63 CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 100 -I - 80 60 - 0 0 I1 E - / 40 0 - 20 0- .1r flat, Q flat, 0 dB HFE, Q Condition Figure 5.10: Keyword identification for impairment PB PB KG 64 CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 65 For the impaired subjects PG and AL the matching of the raw scores was very good across conditions. The normals appeared to derive a slightly greater benefit for high-frequency emphasis. Only one normal subject was simulated with the sentences of impaired listener PB. In this case the normal scored significantly better for all conditions. Error patterns In order to better assess the similarity between the performance of the impaireds and the simulated normals the following correlation measure was proposed and implemented. The answers of the impaired subject were compared with those of the simulated normal, keyword by keyword. Each correct identification was marked as 1, each incorrect as -1. The sequences thus generated were then multiplied, point by point and the results summed. This was equivalent to awarding a 1 if the responses were the same (i.e. both correct or both incorrect) or a -1 for dissimilar responses (one correct and one incorrect). This was done for all sentences in a given listening condition. The maximum score for a 30 sentence set would then be 150 and the minimum -150. If the two subjects answered exactly the same way the maximum score would be obtained'. Tables 5.9 and 5.10 list the correlations thus computed for impaired subjects PG and AL. The analysis was not performed for subject PB since her scores were significantly lower than the simulated normal's and the correlation would have little physical meaning. To determine the significance of the resulting "keyword correlation" (KC), probability theory was applied to the product of the two sequences assuming that they represented independent Bernoulli processes. The possible outcome of each process was a 1 or a -1 and the probability of "success" (i.e. a 1) was taken as the percent of keywords identified correctly by the subject. The random variable k resulting from multiplying two Bernoulli variables with probabilities of success P. and P would have a probability distribution Pk(1) = PaPb (1 - Pa)(1 - Pb) 'An amusing similarity in confusions occured for subject PG and two normals on the sentence "The marsh will freeze when cold enough". All three responded with "The Martian __ went home" CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE,0dB SNR 66 Correlation (KC) DW KG AW 42 68 74 50 33 46 60 88 58 92 76 92 Table 5.9: Correlation between keyword identification; impaired subject PG, 150 keywords per condition Listening Condition flat, Quiet flat, 0 dB SNR HFE, Quiet HFE,0 dB SNR Correlation (KC) DW KG AW 38 64 45 78 64 70 116 110 104 98 60 99 Table 5.10: Correlation between keyword identification; impaired subject AL, 200 keywords per condition CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 67 Pk(-l) = 1 - Pk(1) Summing the values of k over all the trials (i.e. keywords) we obtain the correlation KC under the assumption that the response sequences were uncorrelated (the null hypothesis). We can easily find the expected value and the standard deviation of KC under those assumptions: E(KC) = N(P(1) - Pk(-1)) 0 KC = 4x/K(P(1)P(-1)), where N is the number of keywords. By comparing the actual correlation with that calculated under these assumptions we can test the null hypothesis of no correlation. For all cases tested the actual response correlation between the impaired subject and the simulated normals lay at least 3 standard deviations above the mean predicted by the null hypothesis. The hypothesis was thus always rejected at the 0.01 level of significance. We can thus conclude that the error patterns were significantly correlated. 5.3.3 Simulation without recruitment As explained above, the normal subjects were also tested with the same sets of consonant-vowel syllables simulating only the threshold shift of the impaired subject (i.e. simple filtering). This was done to investigate the importance of including recruitment in the simulation. Figures 5.11- 5.13 compare the scores of the impaired subjects with those of the normals without recruitment. The normal scores represent an average of three runs. The simulated normals score consistently better under these conditions than the impaired subjects. The increase in their scores is particularly evident for the flat characteristic where their performance improves by an average of 15 percent. The improvement in the normal subject performance for the HFE condition is less dramatic although still evident (particularly for impairment PG). CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 68 - I n80a) 0 4 60- 0 C.) 40a) 0~ E PG DW KG AW 200 flat, Q flat, OdB HFE, Q HFE, OdB Condition Figure 5.11: Consonant identification scores for impairment PG; recruitment not simulated, 94 dB SPL overall level I n I 80- 0 60- 0 Ul 0j 40- AL DW KG AW - 20 0 flat, 0 flat, OdB HFE, 0 HFE, OdB Condition Figure 5.12: Consonant identification scores for impairment AL; recruitment not simulated, 100 dB SPL overall level CHAPTER 5. DETAILED EVALUATION OF AGC ALGORITHM 179f i nn 800 60- C 40- E3 PB DW KG 20- - 0 flat, 0 flat, OdB HFE, Q HFE, OdB Condition Figure 5.13: Consonant identification scores for impairment PB; recruitment not simulated, 104 dB SPL overall level Chapter 6 Discussion 6.1 Evaluation of performance The data show several trends reflecting on the accuracy with which the AGC simulation reproduces the effects of sensorineural hearing impairment on speech intelligibility. For virtually all subjects and stimuli, when the flat frequency-gain characteristic was employed the results obtained from the simulated normals closely match those of the impaired listeners. This holds true for the consonant reception scores obtained both in the pilot study and in the experiments of Chapter 5. Collectively these studies cover a wide range of hearing loss configurations and degrees, from mild to severe. The scores remain well matched under various signal-to-noise ratios. Similarly, the raw scores obtained by the impaireds during sentence comprehension tests were well predicted by the scores of the normals listening to the sentences passed through the simulation. Tests designed to investigate the similarities between the patterns of perceptual errors made by the impaireds and the normals confirmed these results. We evaluated the perception of consonant features in the CV syllables. In particular, we compared the percent of information available in each feature that was transmitted to the listeners. It was shown that the transmission of feature information by the impaireds was highly correlated with the transmission by the simulated normals. Similarly, the errors made by the subjects in identifying the keywords in sentences were compared. Using the "keyword correlation" coefficient described in the previous chapter, it was shown that the patterns of keyword identification were correlated between subjects with a high level of significance. The algorithm performed less well when the high-frequency emphasis character- 70 CHAPTER 6. DISCUSSION 71 istic was employed. Here, the simulated normals tended to perform significantly better than the impaired subjects. Again, this trend was seen for nearly all stimuli tested. It was also observed that the discrepancies generally increased with the severity of the impairment. This is particularly evident for the three impairments simulated in the foregoing chapter; there are suggestions of similar trends in the pilot study. Here, the scores for subject PG (the least severely impaired) were matched to those of the normals for all conditions including HFE. There were clear differences for subject AL and yet larger differences for subject PB. For each of the last two subjects the discrepancies increased with increasing level of the stimulus. In the pilot study the normals consistently achieved scores above the average of the impaired subjects' scores under HFE. The raw scores obtained on the sentence tests showed patterns similar to those seen with CVs. The discrepancy with HFE was smaller for AL but was quite large for subject PB (although only one normal was used in that particular simulation). The results of the error analyses were less conclusive with respect to the difference between flat and HFE conditions. Feature transmission for the impaireds was about as well matched to the normals under the HFE condition as it was under the flat condition. One notable exception was subject PG whose feature transmission was less correlated with that of the normals for HFE than for flat. Likewise, the error patterns on the keyword identification task generally remained well correlated under the HFE condition despite the differences in overall performance. The interpretation we offer for the observed discrepancy with HFE (and the lack of discrepancy with a flat response) is based on the audibility of speech components in different frequency regions. From inspection of band levels of the CV stimuli (24) we expect that for the flat characteristic the stimulus levels were such that speech-spectrum components above about 2 kHz were below the impaired subjects' thresholds. The HFE filter raised those components above threshold while attenuating the low frequencies. In broad terms this pattern of results indicates the presence of significant high frequency suprathreshold effects in sensorineural hearing impairment that are not represented by the simulation. Such effects are not present when the speech signal exceeds the impaired listener's threshold only CHAPTER 6. DISCUSSION 72 at low frequencies (below about 2 kHz). - Greater upward spread of masking in impaired ears can be offered as an explanation for the observed discrepancy with HFE. The direction of the masking effect greater masking of high frequency signals by low-frequency maskers than the reverse - is consistent with the direction of the observed discrepancy between simulated and impaired performance being greater with HFE than with flat responses. Since the simulation does not explicitly account for masking phenomena the simulated normals are not affected while the impaireds are. One would expect this effect to be particularly evident when the high frequency components are above threshold (in quiet) but are accompanied by low-frequency components. For the flat frequency response high-frequency components are less likely to be above quiet threshold, so that high-frequency information is unavailable to both the simulated-normal and the impaired listeners. The discrepancy with the HFE condition might not be as dramatic for sentences because the impaireds are better practiced to utilize the contextual information. The normals who are not used to the quality of speech in the simulation might require a longer training period. The correlation scores are harder to integrate into this rationalization since they appeared almost uniformly good for all conditions. As pointed out above the one place they do fail is in fact for the HFE condition. At this time I do not have a good explanation for this result although it appears to be consistent with results reported elsewhere (3; 22) which show lack of correlation between raw scores and feature error patterns. 6.2 Importance of recruitment The experiments conducted with the recruitment-less simulation seem to unequivocally indicate that recruitment plays some role in diminishing the intelligibility of speech for the hearing impaired. The normal subjects scored better when recruitment was absent than when it was present. This was true of virtually all conditions for which this test was performed. Thus, while the elevated threshold still appears CHAPTER 6. DISCUSSION 73 to play a paramount role in determining the effect of a hearing impairment the phenomenon of recruitment also has to be accounted for in the simulation. 6.3 Conclusions The simulation proposed in this thesis was largely successful in duplicating the speech reception of the hearing impaired for normal listeners. The experiments conducted indicate that the simulation faithfully reproduces intelligibility scores for speech reproduced with a flat frequency response. At those frequencies it appears that threshold shift and loudness recruitment as simulated by the AGC algorithm collectively determine the intelligibility of speech as perceived by impaired subjects. For stimuli with high-frequency emphasis other suprathreshold effects appear to dgrade intelligibility for the impaired listeners. The effect proposed to be the factor is upward spread of masking. The algorithm in its present form does not attempt to simulate this effect. However, the structure of the algorithm may be conceptually extended to account for this phenomenon. In Figure 3.1 the gain computation for each channel is affected only by the signal level in that particular frequency band. It is possible to simulate more complex interactions by allowing the gain in one channel to be influenced by the signal level in other channels. Future work in this direction should consider such ways of conditioning the gain. While the experiments conducted during this study yield fairly consistent results it would be instructive to obtain a better evaluation of the qualitative fidelity of the simulation. An ideal experiment would have a unilaterally impaired subject compare the sensation in the impaired ear with that in the normal ear listening to the simulation. Due to limited contact with patients I was unable to recruit such a subject for this study. If possible in the future such an experiment is likely to yield invaluable additional insight. Bibliography [1] Diane K. Bustamante. Principal Component Amplitude Compression of Speech for the Hearing Impaired, Ph.D. Thesis, Massachusetts Institute of Technology, February 1986. [21 William F. Carver. Loudness Balance Procedures, in Handbook of Clinical Audiology. Jack Katz, ed., The Williams & Wilkins Co., Baltimore, 1978. [3] David A. Fabry and Dianne J. Van Tasell. Masked and Filtered Simulation of Hearing Loss: Effects on Consonant Recognition, Journalof Speech and Hearing Research, 29: 170-178, 1986. [4] Jean-Pierre Gagne and Norman P. Erber. Simulation of Sensorineural Hearing Impairment. Ear and Hearing, 8(4):232-243, 1987. [5] Daniel W. Griffin and Jae S. Lim. Signal Estimation from Modified ShortTime Fourier Transform. IEEE Transactions on Acoustics, Speech and Signal Processing, 32(2):236-243, April 1984. [6] C.S. Hallpike and J.D. Hood. Observations Upon the Neurological Mechanism of the Loudness Recruitment Phenomenon. Acta Oto-laryngologica, 50:472-486, 1959. [7] Rhona P. Hellman and Carol H. Meiselman Prediction of Individual Loudness Exponents From Cross-modality Matching Journalof Speech and Hearing Re- search, 31, 1988 [8] Larry E. Humes. Spectral and Temporal Resolution by the Hearing Impaired. in The Vanderbilt Hearing Aid Report, Gerald A. Studebaker and Fred H. Bess, eds., Monographs in Contemporary Audiology, Upper Darby PA, 1982. [9] IEEE. IEEE Recommended Practicefor Speech Quality Measurements. Techni- cal Report IEEE 297, Institute of Electrical and Electronics Engineers, 1969. 74 BIBLIOGRAPHY 75 [10] Harry Levitt. /it The Transformed up-down methods in psychoacoustics. Journal of the Acoustical Society of America, 49:467-477, 1971. [11] Harry Levitt. Speech Discrimination Ability in the Hearing Impaired: Spectrum Considerations. in The Vanderbilt Hearing Aid Report, Gerald A. Stude- baker and Fred H. Bess, eds., Monographs in Contemporary Audiology, Upper Darby PA, 1982. [12] F. Miskolczy-Fodor. Relation between Loudness and Duration of Tonal Pulses. III. Response in Cases of Abnormal Loudness Function. The Journal of the Acoustical Society of America, 32(4):486-492, 1960. [13] Patrick M. Peterson and Joseph A. Frisbie. An interactive environment for signal processing on a VAX computer. Proceedings of ICA SSP, 1891-1894, April 1987. [14] D.P. Phillips. Stimulus intensity and loudness recruitment: Neural correlates. The Journal of the Acoustical Society of America, 82(1):1-12, 1987. [15] R.W. Schafer and L. R. Rabiner. Design of Digital Filter Banks for Speech Analysis. The Bell System Technical Journal, 50(10):3097-3115, 1971 [16] Bertram Scharf and Mary Florentine. Psychoacoustics of Elementary Sounds. in The Vanderbilt Hearing Aid Report, Gerald A. Studebaker and Fred H. Bess, eds., Monographs in Contemporary Audiology, Upper Darby PA, 1982. [17] Margaret W. Skinner. Speech intelligibility in noise-induced hearing loss: Effects of high-frequency compensation. The Journal of the Acoustical Society of America, 67(1):306-317, 1980. [18] S.S. Stevens. Power-Group Transformations under Glare, Masking, and Recruitment. The Journal of the Acoustical Society of America, 39(4):725-735, 1966. [19] S.S. Stevens and Miguelina Guirao. Loudness functions under inhibition Perception & Psychophysics, 2(10):459-465, 1967. BIBLIOGRAPHY 76 [20] Edgar Villchur. Simulation of the effect of recruitment on loudness relationships in speech. The Journal of the Acoustical Society of America, 56(5):1601- 1611, 1974. [21] Edgar Villchur. Electronic models to simulate the effect of sensory distortions on speech perception by the deaf. The Journal of the Acoustical Society of America, 62(3):665-674, 1977. [22] Brian E. Walden, Daniel M. Schwartz, Allen A. Montgomery and Robert A. Prosek. A Comparison of the Effects of Hearing Impairment and Acoustic Filtering on Consonant Recognition. The Journal of Speech and Hearing Research, 46: 32-43, 1981. [23] Marilyn D. Wang and Robert C. Bilger. Consonants confusions in noise: a study of perceptual features. The Journal of the Acoustical Society of America, 54(5):1248-1266, 1973. [24] Patrick M. Zurek and Lorraine A. Delhorne. Consonant reception in noise by listeners with mild and moderate sensorineural hearing impairment. Journal of the Acoustical Society of America, 82(5):1548 -1559, 1987. The Appendix A Simulation Program A.1 Full code This is the program used to implement the AGC simulation described in Chapter 3. It was written in VAX C programming language for the VAX VMS environment on a VAX 11/750 computer. It is essentially self-contained except for some I/O routines which were written in the laboratory. The AGC simulation program /* Uses the BNIO I/0 package written by Patrick Peterson. /* All binary input assumed to have the CBG format one block header. */ /* Arguments: 1) input signal file name /* 2) impaired thresholds file name 3) overall output level in dB SPL 4) output signal file name 5) RMS of the input signal /* The following directory is to be defined externally as a logical */ /* name: fildir: location of bandpass filters; filter impulse responses */ /* stored under f"x".s where "x" is the number of the filter. /* /* /* /* /* /* Input and output signals are stored as short integers. Filters are stored as double floats. The threshold file is ASCII list of three rows listing for all 14 frequency bands: 1) impaired threshold, 2) threshold of impairment, 3) impaired threshold of discomfort. This version of the program computes the recruitment thresholds therefore line 2 is ignored #include #include #define #define #define #define #define #define #define #define #define stdio math STMAI BYTBLK SHRBLK INTBLK FLTBLK WINDL DBOUT NUMCHAN LN10 100000 512 256 128 128 /* 20 -200.0 14 2.302585093 window length in milliseconds */ 77 */ */ */ */ */ */ APPENDIX A. SIMULATION PROGRAM #define PI 3.141592 654 #define TINY le-32 #define FILTDELAY 113 /* 78 offset in filterbank to give zero delay */ #define max(A, B) #define min(A, B) ((A) > (B) ? (A) ((A) < (B) ? (A) #define cadd(z, x, y) #define csub(z, X, y) #define cmul(z, x, y) z.real = x.real + y.real; z.imag = x.imag + y.imag z.real = x.real (B)) (B)) - y.real; z.imag = x.imag - y.imag z.real = x.real * y.real - x.imag * y.imag;\ z.imag = x.real * y.imag + x.imag * y.real /* Normal (ANSI) thresholds in dB SPL */ double nthlist[NUMCHAN] = { 45., 35., 23., 16., 11., 10., 10., 10., 10., 10. }; struct 9., 7., 8., 9., cmplx{ double double } real; imag; cxaddo, cxsubo, cxmul(); static int dummy = 0; main(argc, argv) int argc; char { *argv[]; char char int int int short float double double double struct FILE filtname[100], thfile[100], fout[100]; thlist[512], *thlistptr; i, j, fn, siglen, filtlen, mf, ms, wlen; filthead[INTBLK], buf[INTBLK]; floattocmplxo, inttocmplxo, padlen; sig[STMAI], icvout[STMAI]; filt[8096]; env[STMAIJ, expenv[STMAX], fcvout[STMAI], chan[STMAI]; rmsout, impth, recth, dcmth, normth, normdth; recompo, gettho, atof(); cmplx filtcx[STMAI], sigcx[STMAI; *fopen(; for(j = 0; j< STMAI; *(fcvout + j++) = 0) /* initialize sum /* read threshold recruitment threshold and discomfort threshold file fn = opentargv[2], 0); thlistptr = thlist; while ((j = read(fn, thlistptr++, 1)) > 0); close(fn); /* read the input signal fn = openbn(vmstring(argv[1]), &1); getbn(&fn, buf, k(BYTBLK)); */ /* read header */ APPENDIX A. SIMULATION PROGRAM 79 siglen = buf[0]; wlen = (1000 * WINDL) / buf[1]; for (i = 0; i <= (siglen / SHRBLK); i++) /* read signal */ getbn(kfn, Big + i * SHRBLK, k(BYTBLK)); closbn(kfn); /* read filter for channel 1 mkfilt(filtname, 1); fn = openbn(vmstring(filtname), 1W); getbn(fn, filthead, A(BYTBLK)); filtlen = filthead[O]; closbn(kfn); */ /* read header */ /* find FFT of input signal ms = inttocmplx (sig. siglen, (filtlen + siglen), sigcx); padlen = pwr(2, ms); fft (sigcx, ms); /* Loop to process all channels for(j = 1; j <= NUMCHAN; /* j++){ */ read filter, find its FFT mkfilt(filtname, j); fn = openbn(vmstring(filtname), 14); getbn(kfn, filthead, k(BYTBLK)); /* read header */ for (i = 0; i <= (filtlen / FLTBLK); i++) getbn(&fn, filt + i * FLTBLK, k(BYTBLK)); closbn(kfn); mf = floattocmplx (filt, filtlen, padlen, filtcx); fft (filtcx, mf); /* multiply signal and filter FFTs, find channel output for (i = 0; i < padlen; i++) filtcx[i] = cxmul(filtcx[i], sigcx[i]); ifft(filtcx, mf); for (i = 0; i < siglen; i++) chan[i] = filtcx[i + FILTDELAY].real; /* get thresholds: impaired, threshold of recruitment, threshold of /* discomfort, normal and normal discomfort, adjust for external gain rmsout = 20 * log10( atof(argv[5])); impth = getth(thlist, j, 1) - atof(argv[3]) + rmsout; recth = recomp(getth(thlist, j, 1), *(nthlist+j-1)) - atof(argv[3]) + rmsout; dcmth = getth(thlist, j, 3) - atof(argv[3]) + rmsout; normth = *(nthlist+j-1) - atof(argv[3]) + rmsout; normdth = 120. - atof(argv[3]) + rmsout; /* find channel level, gain and processed channel output, add to total log-rms-env(chan, env, siglen, wlen); */ */ APPENDIX A. SIMULATION PROGRAM 80 envexpand(env, expenv, impth, recth, dcmth, normth, normdth, siglen); chanout(chan, fcvout, env, expenv, siglen); } for(i = 0; i < siglen; i++) *(icvout+i) = *(fcvout+i); /* /* convert to int */ write the simulation output fn = creabn(vmstring(argv[4]), &512, &0); putbn(fn, buf, &(BYTBLK)); /* write header */ for (i = 0; i <= (siglen / SHRBLK); i++) /* write file */ putbn(kfn, icvout + i * SHRBLK, k(BYTBLK)); closbn(kfn); /* subroutine to find the name of current filter file mkfilt (filtname, chnum) char *filtname; int chnum; char char char chord[100], *strcato; pcl[100] = "fildir:f"; pc2[100] = ".s"; *filtname = "\0 sprintf (chord, "*d", chnum); strcpy(filtname, strcat(pcl, strcat(chord, pc2))); /* subroutine to get nth element out of the line of threshold list /* indicated by code double getth(tlist, n, code) char *tlist; int n, code; int double i, j; atof(); for(j = 1; j < (code - 1) * NUMCHAN + n; for(; *tlist==' I *tlist=='\n' for(; *tlist >= '0' &k j++){ 11 *tlist=='\t'; tlist++) *tlist <= '9' 11 *tlist== '.; tlist++) return(atof(tlist)); /* subroutine to compute the level waveform of the channel logrms-env(sig, env, siglen, wlen) APPENDIX A. SIMULATION PROGRAM short double { 81 *sig, siglen, wlen; *env; int double i; avs = 0.; for(i = 0; i <= wlen/2; i++, sig++) avs = avs + (double)(*sig * *sig) / wlen; = 10. * loglO(avs); for(i = wlen/2+1; i < ulen; i++, sig++){ avs = avs + (double)(*sig * *sig) / wlen; *(env++) = 10. * log1O(avs); } for(i = wlen; i < siglen; i++, sig++){ avs = avs + (double) (*sig * *sig - *(sig-wlen) * *(sig-wlen)) / *(env++) *(env++) = 10. * loglO(avs); } for(i = wlen/2; i > 0; i--){ avs = avs - (double)(*(sig-i) * *(sig-i)) *(env++) = 10. * loglO(avs); / wlen; } } /* subroutine to expand the envelope according to the recruitment /* function for gain computation envexpand(env, expenv, ith, rth, dth, nth, ndth, siglen) double *env, *expenv, ith, rth, dth, nth, ndth; int siglen; { int double i; lth, pth, eli, s12, s13, icti, ict2, ict3; lth = ith - 5.0; pth = dth - 5.0; /*lower thresh for rise to thresh*/ /*thresh for rise to discomfort */ eli = (nth s12 = (rth s13 = (ndth icti = DBOUT ict2 = nth - DBOUT) / 5.0; /* slope for nth) / (rth - ith);/* slope for pth) / 5.0; /* slope for - sli * lth; /* intercept rise up to thresh recruitment section rise to discomfort for rise to thresh /* intercept for recruit curve ict3 = pth - s13 * pth; s12 * ith; */ */ */ */ */ /* intercept, rise to discomfort */ for(i = 0; i < siglen; i++) if (*(env+i) <= lth) *(expenv+i) = DBOUT; else if (*(env+i) <= ith) *(expenv+i) = icti + eli * *(env+i); else if (*(env+i) <= rth) *(expenv+i) = ict2 + s12 * *(env+i); else if (*(env+i) <= pth) *(expenv+i) = *(env+i); else if (*(env+i) <= dth) APPENDIX A. SIMULATION PROGRAM else 82 *(expenv+i) = ict3 + s13 * *(env+i); *(expenv+i) = *(env+i) + ndth - dth; } /* subroutine to compute gain for the channel, /* and maintain a running output total scale the channel input */ */ chan-out(sigin, sumout, env, expenv, siglen) int siglen; double *sigin, *sumout, *env, *expenv; { int i; *(sumout+i) += *(sigin+i) * exp( LN10 (*(expenv+i) * for(i = 0; i < siglen; i++) - *(env+i)) /* subroutine to compute the treshold of recruitment double recomp(impth, normth) double / 20.); */ impth, normth; { double recangle, slope; recangle = (PI / 180.) * (47. + 0.45 * (impth - normth)); slope = tan(recangle); return(min(max(((slope * (impth - normth)) / (slope (impth + 15.)), (impth + 40.))); } /* conversion of float array to array of cmplx structures int floattocmplx (arr, arrlen, maxlen, arrcx) float *arr; arrlen, maxlen cmplx arrcx[l; int struct { int i, j, padlen; i = 0; while ((padlen = pwr(2,i++)) < maxlen); for (j = 0; j < arrlen; j++){ arrcx[j].real = *(arr + j; arrcx[j].imag = 0; } for (j = arrlen; j < padlen; j++) arrcx[j].real = arrcx[j].imag = 0; return(i-1); /* conversion of integer array to array of cmplx int inttocmplx (arr, arrlen, maxlen, arrcx) *arr; short structures - 1) + normth), APPENDIX A. SIMULATION PROGRAM int struct { 83 arrlen, maxlen cmplx arrcx[j; int i = 0; i, j, padlen; while ((padlen = pwr(2,i++)) < maxlen); for (j = 0; j < arrlen; j++){ arrcx[j].real = *(arr + j); arrcx[j].imag = 0; } for (j = arrlen; j < padlen; j++) arrcx[j].real = arrcx[j].imag = 0; return(i-1); /* decimation in time FFT fft (x, m) struct cmplx x[]; int { m; struct struct int cmplx cxaddo, cxsubo), cxmul(); cmplx wn[STMAX/2], ti, t2; i, j, k, len, gp, gph, gpnum; len = pwr(2,m); for (i = 0 i < len / 2; i++){ wnli].real = cos((double) (2 * i * (PI / len))) wn[i].imag = -sin((double) (2 * i * (PI / len)); } for (i = 0; i < m-1; i++){ printf("%d ",i); gpnum = pwr(2,i); gp = pwr(2, m) / pwr(2, i); gph = gp / 2; for (j = 0; j < gpnum; j++){ for (k = 0; k < gph; k++){ cadd(ti, x[j*gp+k], x[j*gp+gph+k]); csub(t2, x[j*gp+k], x[j*gp+gph+k]); cmul(x[j*gp+gph+k], t2, wn[k*gpnum]); x[j*gp+k] = ti; } } } for (i = 0; i < len; i += 2){ ti = cxadd(x[i] x[i+1]); x[i+1] = cxsub(x[i], x[i+1]); x[i] = ti; } } */ APPENDIX A. SIMULATION PROGRAM /* decimation in frequency IFFT ifft (x, m) struct cmplx x[]; int m; { struct struct int cmplx cxaddo, cxsubo, cxmul(); cmplx wn[STMAX/2], ti, t2; i, j, k, len, gp, gph, gpnium; len = pwr(2,m); for (i = 0 i < len / 2; i++){ wn[i].real = cos((double) (2 * i * (PI / len))); wn[i].imag = sin((double) (2 * i * (PI / len))); } for (i = 0; i < len; i += 2){ ti = cxadd(x[i], x[i+1]); x[i+1] = cxsub(x[i], x[i+1]); x[i] = ti; } for (i = 1; i < m; i++){ gpnum = pwr(2, m-(1+i)); gp = pwr(2, m) / gpnum; gph = gp / 2; for (j = 0; j < gpnum; j++){ for (k = 0; k < gph; k++){ cmul(tl, x[j*gp+gph+k], wn[k*gpnum]); cadd(t2, x[j*gp+k], ti); csub(x[j*gp+gph+k], x[j*gp+k], ti); x[j*gp+k] = t2; } } } for (i = 0; i < len; x[i].real /= len, i++); /* miscellaneous functions: complex add, subtract and multiply struct cmplx cxadd(x, y) struct cmplx x, y; { struct cmplx z; z.real = x.real + y.real; z.imag = x.imag + y.imag; return(z); struct struct { cmplx cmplx cxsub(x, y) x, y; struct cmplx z; 84 APPENDIX A. SIMULATION PROGRAM 85 z.real = x.real - y.real; z.imag = x.imag - y.imag; return(z); } struct struct { cmplx cmplx cxmul(x, x, y; struct cmplx y) z; z.real = x.real * y.real - x.imag * y.imag; z.imag = x.real * y.imag + x.imag * y.real; return(z); } /* integer exponentiation function int pwr(base, expo) int base, expo; { int a; a = 1; if (expo < O){ fprintf(stderr, "pwr: exponent out of range\n"); exit(1); } else{ for (; expo > 0; a *= base, expo--); return (a); } A.2 } Program changes for no recruitment The main alteration necessary to simulate normal growth of loudness above threshold was the repalcement of the env-expando) routine as shown below. In addition, the input had to be scaled up digitally in order to avoid serious quantization problems (the simulated stimuli had much lower amplitude than the original utterances). The scale factor was input as the sixth argument. It was compensated for by appropriate adjustment of the external analog gain. \* Altered level expansion for no recruitment envexpand(env, expenv, ith, nth, siglen) double *env, *expenv, ith, nth; int siglen; APPENDIX A. SIMULATION PROGRAM { int double i; lth, sli, icti; lth = ith - sli = (nth 5.0; /* love r thresh for smooth rise*/ /* to threshold - DBOUT) / 5.0; icti = DBOUT - sli * lth; /* slope for rise up to thresh */ /* inte rcept, rise to thresh */ for(i = 0; i < siglen; i++) if (*(env+i) <= lth) *(expenv+i) = DBOUT; else if (*(env+i) <= ith) else I *(expenv+i) = icti + sli * *(env+i); *(expenv+i) = *(env+i) - (ith - nth); 86 Appendix B Consonant Confusion Matrices The following are the consonant confusion matrices obtained from the CV identification tests described in Chapter 5. All the simulated normal responses have been pooled together to render one matrix for each impairment and listening condition. The impaired subject data for each listening condition appear at the top of each page and the simulated normal data for that impairment and condition at the bottom. The heading of the impaired subject matrix identifies the impairment, overall level of the stimuli (in dB SPL), the frequency-gain characteristic and the signal-to-noise ratio. 87 APPENDIX B. CONSONANT CONFUSION MATRICES 1 17 - - - -1 - 1 - - 1 7 - -- 2 - 3 6 3 4 6 1 - - - 3 2 11 - - 1 2 - 16 - - - 1 - - - 12 - 7 40 33 16 - 1 -S 25 17 - - - - 4 - - - - - 4 - - - 2 - 1 - 17 12 13 13 - 3 3 - 20 11 - - 4 - - - 43 - - 2 5 7 16 - - 17 10 - 4 - 7 11 3 - 3 --17 - 2 2 1 N - - 8 3 - - - - 2 - 18 7 29 189 62 - 4 19 22 41 1 - 3 6 60 106 N 1 1 3 4 2 - 1 - 12 3 - 18 - - - -- - - 1 - 7 1 J - - CH - - ZN - 1 2 - 1 1 I - - - Z - IN -2 35 16 24 1 23 - - - - 34 1 - 1 - - 10 4 R 22 - - - - 6 29 1 - - - - 5 2 - - - - S 13 21 69 63 63 H WE - 2 7 - 8 1 1 - 8 - 4 - 2 - 5 2 - - - V L - - - - - 1 - - - - - - - - 1 - - - 2 - 2 - V - SH 32 432 1080 Presentations - - - 11 - - 6 - - - - S - - TH - - F - - G - - 4 - I 1 9 3 - - - 2 - - - - 106 - - - - 7 3 32 - - - 34 - - 27 - 1 1 1 - - 1 2 19 - - 40 9 64 66 64 34 42 1 45 45 45 4 10 - 2 2 45 5 4 45 45 45 45 - - - Y - 17 - 1 - 9 - 39 -1 - - 15 - 1 - 28 - - 2 1 - - 17 - 1 - - - - - - - - 1 - - - - 1 - 2 1 - 1 3 - - 6 - 33 7 10 37 25 13 2 1 - 10 - - 6 2 -- 1 - D - 4 - -1 - 40.5 Percent Correct out of B 5 -4 - 2 - 17 I 18 - - - 1 - - 6 - - - - VB - -- - - R V L y H 16 - 1 - ZH CH J 13 1 - - 26 - P T I B D G F TH S SH V IN - - 22 - - Three simulated normals T - 11 - - - P 2 1 2 - 3 24 - - - R V L y H WI 11 1 - - 2 1 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - - - - - WE - 2 - - - - - - - 2 -2 - H - - - - - Y - 8 1 2 L - - V - 1 R - 1 5 - - -1 I - 2 N - - 2 J 1 - - - 3 2 1 10 8 3 - CH - 1 12 ZH - - - - Z - - - 6 1 4 10 2 I - - S 2 1 IN - 3 V - - - SH - - - S - - TH - F - 5 1 12 G - 13 - D - 6 1 2 B - I - T 432 Presentations - P T I B D G F TH S SH V IN Z ZH CH J P 47.5 Percent Correct out of - Quiet Impairment PG, 94 dB, flat, 88 6 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 1080 APPENDIX B. CONSONANT CONFUSION MATRICES - 1 1 3 - 1 - 1 1 - - - 2 - 15 25 25 2 4 37 - 2 4 3 1 18 25 49 13 2 3 - - 4 - - - - - 2 - - 3 - - 3 1 - - 1 - - 2 17 38 o B D P T 6 6 2 8 2 - 1 I 4 2 25 1 B 4 1 D - - - G 2 2 4 2 2 1 1 2 - - ZH CH - 13 - - J - 3 - M if R U 1 - - L - - 3 1 - - 36 50 8 13 13 14 6 - 27 - 20 3 2 - - 3 4 - 2 4 6 - 2 1 1 1 - 1 1 2 8 3 10 1 - 1 3 12 18 14 12 15 V - 6 - - - - - 7 I - 4 2 - 4 4 - 3 IN Z 2 - - - - - - 1 1 1 2 2 - - 8 - - 2 - 1 - 5 2 - - 1 - - - 5 - 1 - - 2 30 35 - - 2 - 4 - 19 18 - - - 5 - 119 - - - 3 2 - 2 - 4 55 57 56 40 - 3 - - 6 1 10 - - 1 - - 12 5 45 92 39 19 44 35 - - - 2 3 3 1 - - 2 4 12 12 1 - -1 -1 11 4 2 - - - 3 26 R S - 1 N 10 1 - N CH - 5 -1 1 1 1 3 J ZE - 1 1 - - - 5 - 27 31 11 432 1080 Presentations - 6 3 1 - 2 8 - 1 - 3 SH 43 - - 3 1 8 1 1 1 -I - 23 162 - 1 - 2 -1 10 - 9 2 1 - 2 2 1 7 2 S - 15 - TH - 1 9 1 5 - 1 - y H WN - F - - - - z G - F TH S SH V IN - - K - T 4 15 1 34.5 Percent Correct out of P - - 1 - 16 - 2 2 Three simulated normals 16 - 2 - - - - -I -1 2 5 1 - 1 20 - V L Y I WE - 3 3 - 2 2 4 3 2 1 1 5 - - - 1 - - 1 1 9 35 - 28 2 3 3 4 24 - 22 92 48 1 3 - 1 - - - - - 5 6 1 5 6 - - 2 - 6 3 3 - 1 3 4 - 5 36 4 4 5 13 - 1 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 45 45 45 45 45 45 45 45 45 45 45 45 1 1 45 - - 2 - - 45 - - - - 2 1 1 - - - -1 45 2 - - - 6 2 1 - - 1 13 45 45 45 45 45 1 11 45 45 45 - 45 55 46 - - - 3 WE - - H Y - 1 - L - - -1 -1 V - 4 3 - 6 - R - 1 1 1 2 I - - N - - 2 - 2 - - - I 3 J - - - 1 CH - 1 8 ZN - 1 - 6 1 2 - - - 4 2 2 3 1 - Z - 1 1 1 - - IN - - 1 V - - 2 - - 1 12 SH - - - 2 9 - S - 2 1 2 - 1 TH - - 5 F - G 432 Presentations 31.7 Percent Correct out of - D - B - - - R V L y N WE I - P T K B D G F TH S SH V IN ZE ZN CH J T 0 dB SIR - P PG, 94 dB, flat, - - Impairment 89 39 1080 90 APPENDIX B. CONSONANT CONFUSION MATRICES - 21 4 10 23 z ZH CH - - - - -45 - - -- -223 1 - - - 4 - - - 6 1 - - - 1 25 - - 3 - - - - - - - - - 17 - 5 34 6 29 ---- 6 - 1 1 2 - - - 11 2 - - 5 - 17 - 12 - 43 - - 3 IN - - - - N I - - - - R - - - V L - - - - 3 - - - - - - - 7--- 1 - - - - - - - - - - - - - 12 - - I I - - - - - - 24 47 37 88 47 60 69 22 12 46 2 27 Z ZN CH - - - 1 - 2 - 14 10 - S 6 - 17 11 - 16 - 4 33 5 - J 15 N - - - 1 1 12 - - - 11 30 19 13 13 13 12 13 4 1 - - 3 1 42 - 37 - 1 - 43 30 N 22 L Y H R V - - - - - 1 - - - - - 2 - WE 45 45 45 45 45 - 1 - - - - - 23 -- -- -- - - -- - -2- -5 1- -0 - 4 - -- - --7 - - -8 - 3 5 3 1 45 45 45 45 - 15-- 98 -- - 45 45 -7 -- -- -- - 8 -2 - -6 - --- - - - 46 1 - -19 - -- 4 -3-3 8 45 45 45 45 45 - - - 1 - - 37 - 33 - - - - - 16 45 45 33 25 1080 3 45 - 5 - 45 - -5 - - - 84 - - - 1 - - 45 -- --29 - -102 -38 - - 8 61 432 1080 Presentations 2 4 6 1 45 - J. y H WE 1 - - - 9 2 8 - - 3 2 - 21 - - - 3 3 29 - V - SH - 1 1 1 12 2 1 13 - S - F TH S SH V IN TH - B D G 7 - - -- - -1- 8 - K F 19 - 12 3 - 58.6 Percent Correct out of G 15 1 -1 1 - - 48 D 31 9 - - -1 -8 B 16 - - - 2 -- 7 - P T I T - 1 Three simulated normals P 16 - - - 18 - 1 - -- 4 3 - 2 - 4 - -- 27 - - 18 - - - 26 4 8 1 1 - -- 21 - - - 10 1 - 20 4 - - 4 - - 11 - 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - - - -- - - -1 1 1 2------------1 2- WV - - R v L y H WH - - - - 8 - - - - - 1 1 H - - - - Y - 1 1 L - 1 - 2 V - - - - R I - 2 - - - N - 5 J -I - 14 12 9 CH - - ZN - - Z - 3 15 IN - 1 - V - - - - SH - 3 - - 4 18 S - - - 13 - TH - - - - - F - - - - 5 16 G 432 Presentations - 17 - D - - - - B - - 9 I - - - P T I B D G F TH S SH V IN z ZN CH J N I T 58.1 Percent Correct out of - P HFE, Quiet - Impairment PG, 94 dB, 45 45 - - - - - 29 - - 30 54 29 20 66 46 67 APPENDIX B. CONSONANT CONFUSION MATRICES - - - 2 - 2 - - - - 1 - - - - 1 1 - - - 1 3 - - 5 1 9 27 - 5 32 5 19 K B D G F TH S SH V IN Z ZI 24 - 31 17 8 13 12 - - 2 35 2 3 19 8 3 16 27 - - - - - - - - - 9 - - - - 2 11 - - - I - 1 F TH S SH V II Z ZH CH 2 2 12 - - 3 - 1 - 12 - I - 1 - 15 5 - 1 1 2 1 1 1 15 - 1 6 1 1 5 19 - - 1 - 42 - 18 5 - 6 - - - 1 4 27 - - 2 - - 3 5 1 36 1 29 1 - - - 1 - - - - - - - - - - - - L - - - 15-----------------5 ------ - --1- - --3 2 10 - - - y H WV 1 - 1 - - - - - - 0 70 R W - 33 - - - - - 6335 - - - -- 4- ---- 5 - 6 1 63 - - - 5 - - - - - - 1 40 34 - 5 1 81 9 1 - 12 - 14 21 12 15 - - 1 29 10 24 11 17 52.3 Percent Correct out of T B D G 6 - -6 1 - 4 12 15 - 13 - - - - - - - - - - - 7 - - - 28 23 49 I 55 CH 13 11 - K - - 1 P - P T 8 2- - 18 Three simulated normals ----5 1 3 - J 45 31 - - - R V L Y H WE - 2 1 1----------------3 2 2 - 8 3 - 1 5 - 45 45 45 45 45 45 - -3 1 8 -- - 4 - - - - 1 - 7 - - - - - --- - -1 - - - - - - - 5 - 1 - - - 1 - - - - - 35 100 432 1080 Presentations I N 3 - 40 5 - 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - 22 - 2 2 -- -1 -4 1 - -- 2- --- - 24 3 13 - 20 -- - - 31 -0 2 14 16 12 - - -1 - - - - - - -7 - - - - - 14 - 3 - -I - - 1 - 1 1 - - 1 - 1- - -- - 2 - - - 1 -- - - - 3 1 4 1 -3 - 5 - 45 45 45 45 45 45 45 1 - - - 45 45 45 45 45 - 29 - 23 3 45 2 - 20 - 2 - 10 45 45 45 50 41 30 56 45 61 43 19 1080 - 26 - 45 45 45 - - - - 2 - - - - 3 - - 1 -1 -- - 1 - 3 - - 1 - 3 - - - - - - - - - - - H WE - 2 4 1 - 3 - - 1 7 6 13 - 4 1- - - 6 - - Y - - 1 L - 6 10 - V -I - - 1 R - - - 1 9 I - - - - K - - 13 - J - - 18 2 - - - - 2 2 2 - - CH - 1 1 - 1 - - - ZE - - - - - Z - 1 - - - 10 - IN - - 16 V - - - 2 SH - - 17 - S - 7 - TH - F - G - D - - 2 B - 5 I - P T K B D G F TH S SH V IN z ZI CH J N N R W L y H WE T - P 432 Presentations 51.9 Percent Correct out of - Impairment PG, 94 dB, IFE, 0 dB SIR 91 5 APPENDIX B. CONSONANT CONFUSION MATRICES Impairment AL, 94 dB, flat, Quiet - 24 7 7 - - 26 - 15 2 - 1- 2 - - - - 1 - - - 3 1 - - 43 12 - - - - 5 6 - - - 1 1 -- - 3S 61 - - 0 - - 46 1 9 - - 17 - - - - 26 - - 45 10 32 - 5 - - - - - - - - - - - 4 - - - F TH S SH V S - - - 5 - 5 10 1 25 13 - - - 15 - 2 - - - - 6 7-- - 2 - 4-5-- 14 116 29 107 - - - 5 - - -- - 73 - 1 11 1 - 3--------- - 6 2 - 5 - - - - - - - 4 - - - - -1 - - 1 - - - - 2 5 - 4- 3 3 2 - - 14 3 2 1 - 9 2 3 20 5 -15 4 7 12 648 - - 2 -24 - - - - - - 42 3 -- - - - - - ---- - - 6 - - - - 22 -- - - 25 13 22 4 18 49 2 - 6 0 IN 2 10 1 - 1 5 - 9 5- 3 53 - - - - Z ZH CH J 7 - 21 4 1 23 2 - 26 25 N - - 1 6 2 2 9 29 38 - 20 62 1296 Presentations I R V L Y I WE 1 3 1 - 6 1 - - - - 1- - - 3 - 2 - 1 12 - - - - 8 2 1 1 - - --- -- - - - - - - - - 2 - 1 - - 4 17 - - 34 - 20 - - - - - - - - - - - - - 10 - 45 - 32 - 10 34 - 1 - - 50 - - I - 10 84 14 - 35 88 31 - - 17 1 10 14 - - - 6 4 63 - 98 1 1 1 - 36 7 21 - - - 4 1 -11 - 7 - - 2 4 28 2 8 7 2 10 29 - 26 2 - 11 78 85 60 34 1 3 5 - 9 ----- 58 - - 3 - - - 1----- -----11 51 - - - - 2 - - I--1 46 - 18 10 - - 53 -4 - - 2 -1 - 20 25 - G 10 - - D - 4 - - 2 - - 3 3 - - - - - - - - 19 49 25 -1 - 4 - - 50.0 Percent Correct out of B K - - P T K B D G F TH S SH V IN Z ZH CH J N I R W L y H WE T - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 -1 22 - - - 1 Three simulated normals P S - - - - - -1 26 - - - - - 53 - - - - 2 - 26 1 - - - - - -I - - - wi - - 1 2 I - - 2 Y - - 1 L - - 1 V - - - R 53 - - 1 -1 I - - -- K 18 - - -- - J - - 2 - CH - - 13 - - 1 - - 1 - - ZN - 1 3 - 1 Z - - 16 - - IN - - - - - 5 3 - V - W L y N WE - - 7 27 1 - - XIH ZE ZN CH J N 1 - - 1 - - 1 10 1 - - - - - - 1 - - SH - 9 - S - 1 7 - TH - - 1 - F - 1 3 17 G 648 Presentations - 5 27 7 11 - - P T K B D G F TH S SH D - B - K 54.8 Percent Correct out of - T P 92 9 49 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 1296 APPENDIX B. CONSONANT CONFUSION MATRICES - 13 1 - 13 - - - - - - 6 3 2 5 - 1 - - 1 5 - 2 10 1 - 8 - - - 7 - 41 11 93 0 1 - 51 3 2 - 1 1 2 7 13 8 - - - 25 - - 3 1 - - 9 - - - 40 2 1 - 16 - - 1- 1 - - 2 - 2 31 17 8 47 19 15 12 - 7 4 - -I - 7 IN Z ZN CH J N - -- 7 - 1 - -- -- 1 - -- 7 - 3 - - - - 6 - 48 - 17 - 13 3 - - - - - - 1 - - - - - 5 - 1 - 4 - 1 1 6 - 1 - - 4 - - 22 121 18 97 76 19 62 2 46 52 - - - - - 2 - 3 2 3 4 - 53 - - -- -- - 6 4 - 23 49 22 - 35 - 19 4 - 1 9 16 60 27 - - 1 - - -9 - - - 92 - - - - 4 - 6 - - - 4 - - - 3 7 4 7 - - 29 5 31 18 N R W L Y 2 2 2 - 2 - N WE - 10 2 2 - 7 3 - - - 1 1 - 2 13 11 - 12 - - 1 1 3 3 - 3 - V - - - - 648 1296 Presentations -- SH - 3 1 86 - - 11 16 - - -4 24 20 - I - 10 6 - 11 8 - - 4 - 4 14 - - - - -- - - - -- 8 1 - 3 - - 7 8 14 26 - - 1 - - 22 4 25 1 - 7 5 - - 2 11 91 31 70 3 1 54 54 54 54 54 54 7 3 - 1 1 - 1 - - 2 8 1 5 1 - 1 6 8 - - - - 1 6 5 2 2 2 1 - 6 7 - - -1 7 54 54 54 54 54 54 54 - 6 2 I 8 - - 4 17 2 1 1 - - - 33 35 - 6 - - - - 11 2 S - 12 TH - 14 - 33 - - - - F - G - D - B - ZN CH J N N R W L y H WE I 2 2 - z T 45 25 16 42.7 Percent Correct out of P 3 - - 7 - 2 - 23 1 7 5 - 1 2 - 56 - - 6 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 1 - 1 - - - - - - 3 5 3 Three simulated normals P T I B D G F TH S SH V IN 1 - -1 1 5 - 2 2 20 - - - - 1 - - - 1 1 2 - - - 5 2 14 25 1 2 3 16 34 13 76 - 514 - - 2 - 47 - 14 - 99 63 55 23 1 28 - - - 27 - H WE 3 13 54 54 54 54 54 54 54 - - Y 54 - 8 - - L - -1 12 - - - 1 V - 1 - - 9 - N - 1 N - -2 W L y N WE 9 7 1 J - 3 - CH - - - ZN - - 3 Z - 3 - - 1 8 - IN - - 1 4- 2 - - - - - 10 V - - - 1 1 8 - - - - 1 1 - - - - 4 SH 27 - S - 16 -1 - R - - TH - 2 F - 8 27 8 G - 2 - D - P T I B D G F TH S SH V IN Z ZH CH J H I B - I - T 648 Presentations 39.5 Percent Correct out of - P 0 dB SIR - 94 dB, flat, - AL, Impairment 93 54 6 54 17 54 56 1296 APPENDIX B. CONSONANT CONFUSION MATRICES AL, 94 dB, P T I 18 9 26 - - - - 3 3 4 9 4 - - - - 2 1 - 17 67 16 1 1 - 7 7 80 1 - -- - - ---59 - - 31 - - - 14 R - - 3 310 19 1 - - 4 - - 2 -- 24- 53 1 - - - -6 - - - - 2 - - 9 - - 7 95 35 86 S 2 3 - - - 41 88 38 6 61 1 - - - y H WE - -2 46 - - - 13 - 24 10 -2 -5 2 - 2 - 2 66 - -1 23 - - - 2 8 - 6 - - 4 1 - 3 - 7 21 --- 2 15 i -- 17 29 37 - - 19 52 7 -1 2 7 - 2 3 1 13 7 3 1 2 16 8 20 11 30 -- -1 -2 - 271- - 19 -- 1 -3 7 648 1296 Presentations - CH J N I R V L Y N - - - - - - - 3- - - - - 7 - - - - - - - - - - 54 - - - 76 - - 3 54 54 - - - - - - - 5 118 1 4- 1 - 3- - - - - - - - - - - I 19 - 1 1 - 14 - - 36 14 27 3 - 3 24 4 - - 9 3 37 - -10 10 -51 - - - - - 13 - 2 1 59 - ZN 1 - 55 - Z - 2 - IN - - - - V L - - - - N N1 - 37 ZN CH 2 V - 8 z J SH 1 - 14 S - - - 3 1 28 - TH - 5 - 1 F - 16 54 G - D 19 - 1 2 - 1 - 27 - - 10 1 1 -- 9 1 1 - - 27 - 1 - 4 - 7 - - 6 60.0 Percent Correct out of - B 1 - - - - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 - - 21 1 - 2 - - 1 - - - - - - - - - - 17 27 - - - - - 1 - 1 - - -- 1 - 2 2 - 4 - 2 - 13' - 92 35 54-2923 - - 43 - 94 - 1 - 53 4 2 56 1 - 1 35 6 21 31 32 63 5 68 WE - 5 7 1 6 - - - 23 - 64 37 50 - 54 54 54 54 54 - -2 - - H WE Y 54 - 2 - F TH S SH V IN L - - - 9 - 2 - V - - 19 49 10 - R 1 54 54 54 54 54 54 54 54 - - 2- 26 7 I 54 12 54 54 - - -1 I 7 - N -- - - -- 1 2 - 2 - 8 - - - - 17 - T B D G J - - - 19 4 8 271 - 1 - - - I CN - - 12 - P 1 1 ZE - -- - 14 Three simulated normals P T Z - 56 IN - 21 1 V - 2 - N R L y H WE - - 1 SH - 1 - - S 648 Presentations 1 8 3 - - 8 3 19 27 9 - - - TH 19 7 - 1 F G - 8 D 49.1 Percent Correct out of - - B Quiet - P T I B D G F TN S SH V IH Z ZH CH HFE, - Impairment 94 54 54 1 10 54 24 1296 54 APPENDIX B. CONSONANT CONFUSION MATRICES - 3 1 - - - - - 7 1 - - - 2 - - - - - 1 ZH CH - - 3 - 1 1 - - - - 1 -3 - - - 1 - 2 - - - 1 24 17 - 1 1 2 2 - - - - 27 2 - - 7 2 - - 3 2 -- -- 2 - - - 2 4- 9 13 - 1 27 -3 - 3 11 2 1 - 40 15 28 0 46 31 1 G F 4 2 - - - - - 1 - - 1 4 49 4 2 - - 4 - 9 - 14 - - - 31 2 - - - - - I 1 - 1 -22 22 6-4 - - - - - 88 - - 6 25 75 6 - 77 S SH V IN Z 4 1 3 2 1 - 8 - - 15 - - 11 - 5 10 5 - 4 2 1 - 1 2 8 3 9 1 25 2 1 2 18 - 7 42 98 16 14 -1 - 44 N 3 13 -1 1 4 - - - - - - 3 - 1 - - 10 3 - - 18 - - - - - - - - - - - - 28 - 1 - - 5 I - 47 13 3 7 - 2 12 3 - - - - 7 - 7 - 58 52 19 -I 1 I 67 - - 54- 8 - - - - 3 - - - 1 - - - - -4 1 - - - 10 48 1 - - 6 10 - 1 53 27 - - - 65 - Y 3- 53 33 - 27 103 -5 6 24 17 - 1 - 63 - - I 12 22 12 11 2 7 2 6 648 1296 Presentations L 1-- 5 20 V - 14 32 R - - - 12 I - - 8 K - -- 2 3 J - - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 9 CH ZN - - 1 3 - 35 - 3 5 V - 4 - N N1 5 - TH - 1 20 10 y H WE D 16 50 59 - 14 B 13 - 22 7 - I 10 - T 0 45 53.4 Percent Correct out of -- - - P - L - - - - 1 - 4 - - - - - - 3 - - 14 31 - 13 - 1 - 6 7 24 10 30 75 57 26 68 1 6 1 3 2 3 16 - - - - - 1 - 8 5 WE 54 54 54 54 54 54 1 1 54 54 54 54 54 54 54 54 54 3 - 8 - - 22 -- R - - - - 18 - 3 H WE 2 47 - 24 - - - 61 - F TH S SE V IN z ZE CH Y - Three simulated normals B D G 4 2 3 14 K L 25 y N WE P T 1 -1 - L - - -- - - -- 1 1 - 7 4 - J m I R v - 1 - V R - - - 6 3 - I - 1 9 - N J - 5 - CH - - - - 2 ZN 14 54 54 54 54 69 51 59 35 1296 5 5 31 - 1 1 - - - - - - 2 26 Z - 11 - IN V - 1 SH - - S - 12 4 1 TH - 7 25 14 7 2 F - 10 2 1 1 G - D - - B 648 Presentations Correct out of - Z K 45.2 Percent -- TH S SH v IN T - B D G F P - I 0 dB Sfl - P T HFE, - Impairment AL, 94 dB, 95 1 17 1 54 54 54 54 54 V IN Z ZH CH J N 2 6 12 - - 4 18 2 - 9 - 12 - - - - - - - - - - 6 18 - 9 3 - - - 13 - - - - - - - - - - 18 11 - 3 - - - - --1 - --- -4-- -----------2 - ------- 1 - - 4 - - - 8 - - - - 6 18 - - 1 - 2 2 - 2 - - - - 17 27 18 23 24 16 - 17 2 3 9 2 3 - - - - 10 7 19 24 15 32 16 1 51 0 23 19 20 0 13 9 29 10 20 16 4 - - - 17 - - - 16 - 35 - - - 1 1 3 37 - 48 24 22 4 - - - - 26 - 17 3 1 1 - - - 1 - 2 - - -1 N - 2 - -1 R - - - - - - 3 - - - 31 78 10 1 - 5 - 1 - - 1 7 32 37 2 - - 1 - - - 1 - 3 - 7 33 - - - - - - - 4 - 2 IR Z ZH - - 3 - V 1 - - - I - - - 21 - - I- - - 2 - - - - - - - 3 SH - - - S 52 7 1 2 8 - TH 18 - F 7 - G - y H V 1 62.9 Percent Correct out of ZR CH V L - - D 2 1 -4 - 6- - -51 2 - 42 - CH J - - 53 20 6 - - - - - 11 - - 3 1 - - 8 1 45 122 79 16 35 52 59 1- - 8 18 18 18 18 18 18 3 3 432 I R V L Y R VI 1 - - - - - - 1 - - - - - - - - - - - - - -2 - - - -- - - -- - 2 9 1 31 - - - - - - 54 - - 54 54 54 2 54 - 24 - 73 55 44 - - - - - - - - 75 45 76 43 - 4 18 4 3 1296 Presentations K - - - - - 1 2 43 - 23 11 I 21 64 50 34 69 40 14 29 100 S S - - z - -1 - - B - - - - - IN ZR 4 1 - - I I 1 - - - - 2-- ----16 F TH S SH V - - 2 - D G - - T 52 - - - I B 24 - - - P T - - Three simulated normals P - -6 11 3 9 - 3 43 1 7 11 3 6 43 1 54 54 - - 18 - - 112- - - - - 3 - - - -- - - - - 54 54 54 54 54 54 54 - - - - - 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - - - -I -- - - V N - - - 14 Y 1 S - - 2 - L - 5 V - 18 R I - SH - - S - - - TH - - F - - G - - D - - - - B - 5 - I - - 10 1 - T 432 Presentations - - - P T K B D G F TH S SH V IN Z ZR CH J x I R V L y H VK 62.5 Percent Correct out of - - P 96 - Quiet Impairment AL, 100 dB, flat, - - APPENDIX B. CONSONANT CONFUSION MATRICES 6 14 54 54 54 54 54 54 54 54 54 54 54 1296 APPENDIX B. CONSONANT CONFUSION MATRICES -- -1 -- 13 - - 2 - -- - R - - 4 L y HI - 30 13 20 -- - - - - - - - - - - - 2-3 - - - - 3 25 12 42 0 24 17 24 L y H WE - 3 21 2 27 10 - - I 3 27 2 - - - - - - - - - - - 1 -- - - - - - I - S SH - 19 3 - - - 45 8 19 16 - - - - - 23 21 3 - - 2 - - 1 12 - 16 1 - - - 1 5 9 - - - 38 - - - 54 - - 1 - 3 - - - - - - - - - - 10 - - 6 - - 3 - - - - - - 6 3 3 - - - - - TH - - - 4----------------- - - V 3 - 1 - -- 2 1 - - 1 - - 2 - 1 1 -1 - - - 4 - 3 - 15 - - - - - -- 1 - 10 - - - -12 - 1 9 5 2 - - 14 6 14 3 8 27 1 1 53 2 1 - - - - - 1 1 27 19 13 12 V L Y N WE - - - - - 1 - - - 4 2-- -- -- 2- -- 3 3 --2 3--------------1 - 12- -- 1 1 - 6- -- 1 1 - 3 - - - - -- 3 11 10 1 - - 7 - - 2 1 - - 11 - 8 - - 2 - - - - - 2 27 87 32 95 16 R - 4 - - 1--- - - - - - - 3 - 41 - 10 27 - 53 - 53 26 - 8 - 26 119 I - 1 39 54 1296 Presentations N - - - - 3 - 55 8 68 41 80 - 38 1 3 35 S1134 1 1 4 1 3 82 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 432 J ZH 5 25 CH Z 5 70 - - 13 IN - 3- - - - 53 - 15 3 - 3 - 3 - F - - 18 G - - D - B 14 - 0 - I - -4 - N R T - 10 13 57.4 Percent Correct out of Three simulated normals P T K B D G F TH S SH V IN Z ZH CH - - -2 2-1 -- - - - - 2 4 - 2 39 - 2- 10 - 6 1 - WE - -- 21 - - 4- -8 - 12 1 12I1 -1 1 8 - - - 54 5 7 28 4 3 27 2 - 28 - 37 86 58 60 68 1 30 1 - 6 39 1 1 1 5 - -4 -1 4 - - 1 1 1 4 - - - - - - - - - - 2 - P - - - - -5 - - WE - - 18-- - K - - - 1 17- -4 6 - - Y - - 13 5 L -- - - V - 3 1 2 - - R - - - - I 11 - - - - - - - 1 - - - - - - - - I 2 N - - - - 5 - J 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 10 54 54 - - - 1 3 CH - -1 - - ZN - - - 6 - Z - 2 17 - - - 9 - - IN - 4 V - 1 2 SH - - - - - S - - - 11 TH - 1 F - 3 - G - 1- D - 18 7 - - - - B 432 Presentations - za 10 - - P T I B D G F TH S SH V IN zZN CH K T 52.1 Percent Correct out of 0 dB SIR - P AL, 100 dB, flat, - Impairment 97 54 54 2 10 54 31 1296 64 APPENDIX B. CONSONANT CONFUSION MATRICES Impairment AL, 100 dB, NFE, Quiet SH - - 15 - - 2 1 18 1 17 - 10 14 9 4 3 9- 9 - 73 0 2 34 28 26 - - - - - - - - - 2 2 --21 1 V L y H WE 45 63 -- - 2 1 - 45 55 87 1 18 16 5 - 2 12 1 3 40 8 18 12 6 18 - - - 1 14 - - - - - - - - - - - - - 3 - 7 - - - - - - - 2 - - - 1 2 4 - 3 7 9 - - 1----------5 7 - 1 2 - 7 - - - - - - -- -- -- -- - - 33 - 42 - - - 45 3 - - 24 513 - 1 11 -I - - -- - -- 8 1- 16 8 - - - 13 - 37 107 17 -- - 7 58 54 38 - N R V L Y 4 H 9 - V - -- 1296 Presentations H 2I - 432 11 J - - 17 CH SH -18 10 ZN S 6 - 4 52 1 - -3 R 1 16 - - -51 - 14 - - - - -2 - 18 54 9 - 1 30 - 11 Z TH -- 5 2 - - -I IN F - - - 1 - -1 - - 15 3 - - - - - - - 35 - 8 1 G - 6 - 2 53 D - 1 - - 67.4 Percent Correct out of - 1 B - -- N 34 I T - 1 - Three simulated normals P 6 1 - - 13 76 45 1 1 83 54 54 - -54 -2625 - 50 wi - 39 -I 51 4 - 54 14 - - 1 8 - 39 22 2 3 1 51 - 54 62 - - - - 54 54 - 14 4 - - - 17 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 54 54 54 54 54 5 54 54 54 54 54 40 1- 3 - 30 3 62 45 33 45 12 5 2 - 31 - 83 62 54 - - 25 - - - 12 H WE Y - - 1 - 1 -6 - L 54 54 54 54 12 54 - - - 2 - 2 -1 -3 -1 - V - V L y H WH 4 3 - 1 - 1 - R - 13 I 1 - 2 H - - J - 6 CH - - - 18 6 ZH - 10 - Z - 8 - R P T K B D G F TH S SN V IN Z ZH CH IH V - S - TH - 1 - F 432 Presentations - 6 G - 1 17 - D 52.8 Percent Correct out of - 12 - B - T - P T K B D G F TH S SH V IN Z ZH CH J H N P 98 54 54 54 1 3 16 54 1296 APPENDIX B. CONSONANT CONFUSION MATRICES 2 3 39 4 2 0 32 27 38 - - - - 12 3 - TH S SH - - - ---- - - - - - IH z - ZH CH - 52 12 1 7 ---26 - - - - -- - - - - - - - L y - - 112 - - - H WE 3 - 1 - 12 - 63 40 82 73 R - - 43 - - V IN - - - 9 - - 33 - 2 11 - - - - 2 36 13 - 11 - 42 2 - 8 2 6 - - - 12 - 2 4 - 2 - - - - - N SH - - 1 - - - 13 - - 15---------- - I - - - - - 9 2 - 9 - - 37 1 - - -1 - 3 - 17 1 11 - 11 32 13 20 - 1 89 - - - - 54 - ------- 5 - - 4 11 - - - - Z 42 7 7 11 - V L Y K - - - - - - - - - - - 1 - - - - - - - - -10 - - - - - - - 6 - - 7 - 1 - - 3 I 1 - - - 2 - - - - 4 - - 1 - 5 2 - - 1 - - - 1 1 1 - 1 - 2 - 4 - 2 - 11 50 1 - - - 7 - - - - 3 - - - - 26 44 - 58 - 4 66 WE R - 15 64 - 51 28 - - - 2 45 - - - - - 8 1 21 - - - - - 49 91 1 - 611 39 9 1 1 432 1296 Presentations N 4 6 1 K 2 ZH 38 J - - 1 CH - - - - 2 21 2 1 S 1 2 3 2 3 - - 7 1 - TH - - - F - - - G - 1 1 0 - D 1 1 - - B D G F I B - 1 10 62.7 Percent Correct out of I 29 3 - 11 - - 1 1 - 44 2 - - 25 - 1 - - 2 1 3 45 65 66 30 5 - 1 - 38 2 78 35 - - 5 - - 1 - 52 1 4 3 - - 30 1 7 1 - 2 - P T T - - - - 1 - 10 Three simulated normals P 18 16 - - 12 - -2 - - 23 - - - - 12 - - -1 - 20 - 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - 13 - - 29 - WE - - 6 - 1 4 38 1 - - 12 5 5 67 - 23 - 61 30 54 - - - - 44 12 - I - - L y NK - Y - 4 - - L - 1 - V 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 2 54 54 54 54 14 1296 - - 1 - v R R - - - 9 5 3 - 3 - 1 I 2 1 17 3 - - ZN CH I - z 3 4 N 11 6 - 2 - - J - 1 CH - - - 1 ZN - 1 8 - - - 6 1 Z - S 18 - IN - 2 V - 1 S 3 - 1 4 - 16 - - SH - 3 S - - TH - 3 F 6 3 - G - 1 D - 1 13 B 432 Presentations - 6 2 - I - T - P T I B D G F TH S SH V IN P 45.1 Percent Correct out of - - Impairment AL, 100 dB, NFE, 0 dB SIR 99 APPENDIX B. CONSONANT CONFUSION MATRICES 56 36 2 - 3 2 1 1 - 4 - --- 14 1 - - 4 4 2 33 1 38 47 1 2 - 9 34 -1 CH 2 N - - 2 39 23 15 29 1 4 7 8 3 4 - 1 1 1 1 - - - 8 8 2 1 - 8 2 3 - 1 - 1 - CH J N - - - 1 - 1 2 - 14 2 - 3 - 1 4 3 - - - 5 1 - - - - - 32 3 3 - 1 9- 6- - 10 1 - 11 4 - - - - 1 52 - 2 - 11 -10 - - 5 7 1 1 - 1 - - - - - - - 2 - 3 3 1 - ZN - -7 - 22- 7 1 6--- 1 - 11--- 1 19 2 45 1 - 28 18 ---- ---- - 1 - - 41 9 18 6 14 - I 2 - 12 2 Z 2 7 8 6 34 - 7 - 4 1 54 16 148 44 14 67 14 6 - 84- 53 127 17 - - 3 22 - - 3 5 89 S2 113 - 1 1 4 - 10 - - 4 7 2 4 3 12 2 1 12 2 22 5 1 7 1 26 68 32 62 9 648 1296 Presentations IN 7I 7 - 5 R W L Y H 1 - 9 - 1 1 - 2 54 4 - - - - 7 - S - 1 54 54 - 2 - 3 10 3 - 12 3 - - 1 1 1 7 4 WE 8 7 2 6 54 54 54 - 3 54 54 5 1 3 54 54 54 54 - - - V 2 - 1 1 4 27 3 5 - y H WE 2 1 - L - - 1 - 12 2 - R W - 5 2 20 1 6 - -S 1 - 1 - 3 1 3I - - ZN 33 10 3 - - IH ZE - - 9 SN - 3 13 17 11 9 5 7 - - 1 2 6 - 1 - - 5 1 - 23 F - - - 1 TH S G - 1 9 1 - 6 D - 2 14 - 16 - 32.6 Percent Correct out of - 4 22 10 -1 3 2 10 - - - F TH S SH 13 1 4 - D G B - K B I - P T T -1 - 3 16 1 Three simulated normals P - 4 4 1 2 4 2 1 6 7 - 3 - 8 - 1 - - 3 14 2 4 7 - - - 3 13 49 11 - - - 9 1 4 6 2 -1 - 1 1 - - 2 - 2 3 3 - 1 1 - 28 - - - 3 4 1 - - 22- - 1 - 1 - 2 21 16 - 3 - 1 1 4 3 41 5 - 4 1 2 1 2 - - - 2 1 - 1 - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 1 57 54 54 54 7 3 8 1 5 18 2 - 3 - - 8 22 2 8 15 6 27 84 5 82 1 - 87 - - 3 - - - WE - 5 - K - - 2 7 1 1 1 1 - 4 1 - -I - - - - - 6 2-I - - 9 1 4 - - - 1 6 - 1 1 4 - - - - - - 3 2 2 2 - 3 - -4 -8 - 2 Y - 21 - - S 8 - - L - - 5 - - V - 2 2 R - 3 1 I - S 2 - 1 N - - 6 - 2 5 1 J - - - 4 1 - CH - - 2 2 2 2 ZN - - - 5 Z - - 3 3 IN 648 Presentations - - - V - 1 - - 1 12 2 SH - 7 - S - 2 S - - - TH - 1 3 F - 2 10 G - 9 D - -1 - 1 - - - 1 1 - 1 B - R W L y N WE 2 11 2 I - ZH CH 1 1 - z T - P T I B D G F TH S SH V IN P 29.2 Percent Correct out of - Impairment PB, 98 dB, flat, Quiet 100 1 54 54 54 54 54 54 64 - 9 54 54 49 47 1296 APPENDIX B. CONSONANT CONFUSION MATRICES 62 27 38 36 11 - 2 - 18 5 - - 11 4 - 13 - - 10 10 - - 8 2 8 - - 7 4 7 6 - 4 3 - - 1 - 3 S - 4 - - 9 3 - 10 4 3 - - - 6 2 - - 5 10 2 1 2 4 - 5 11 167 72 12 86 1 1 - 5 4 1 - 39 78 - - - 1 - 5 3 10 31 3 1 38 - 14 7 7 - - 3 - - 4 1 - - - - - - - 13 2 4 S5 15 4 - 3 - 11 3- Z 2 - - I - 8 4 IN - - - - - 2 12 18 14 2 - 9 10 - 2 - 1- 14-2 4 6 2 1 16 23 25 3 - 22 2 4 - 1 1 2 2 26 CH -- 6 - - 3 - 6 - 1 2- 9 42 126 ZN 5 - - - 1 1 - 2 - 5 1 3 2 - 4 - 3 1 4 35 33 77 - - - - 1 2 6 2 1 1 4 - 2 5 - 8 29 20 13 648 1296 Presentations - 6 - 1 1 45 - - 2 33 -6 25 - 11 - 37 3 1 - - 33 - - - - - 4 7 2 - 1 3 2 I 1 - 4 5 5 5 1 N - 2 3 1 I 1 J 9 1 -6 3 - - 1 R V L Y - 1 1 7 - - V - - SN - 1 8 1 - 2 4 - - 1 7 13 -29 9 - - S - 1 1 - - - - - 6 1 - - - 3 14 9 2 1 1 - 2 2 3 - - - - - - 3 6 - - 7 1 2 5 1 2 2 3 - - - 4 - - - 1 - - - 3 - - - 11 2 - 16 11 2 12 9 78 41 43 38 17 55 2 3 2 13 - - I - 6 5 5 12 3 5 - 1 2 1 5 12 14 - 1 4 - W - - 1 N 15 3 15 2 - 9 - 1 - - TN - F 4 23 - 0 5 2 - D - - 1 B - 3 - 1 1 1 1 2 1 I - 5 10 1 1 1 30 7 1 4 23.1 Percent Correct out of - - - 6 - P T K B D G F TH S SH V IN Z ZH CH J N I R W L y H WE T 1 3 4 3 - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 3 12 3 - 2 - 28 Three simulated normals P 6 - 2 2 6 4 - 7 - 1 4 3 - 29 - - - 22 1 - 7 1 1 - 1 3 2 2 - 21 H W 1 - - -6 5 6 3 Y - 2 9 - 9 - - - 2 3 4 - I 3 3 6 L 1 S 1 5 1 - 9 5 6 U 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 - 6 54 54 54 4 47 - 2 10 8 54 54 54 54 - - 10 54 66 47 1298 - 18 18 5 27 10 - 124 102 1 - 1 1 8 - 2 R - - - L y H UN 2 17 I 1 - 1 2 1 -1 1 2 8 3 I 1 1 1 R W 1 N - -1 - I J 2 - 1 ZN CH J N CH - z 2 2 2 ZN - - Z - 1 IN - I - V - - SH - P T K B D G F TH S SH V IH S - TH - F - G - D - B - I - T - P 648 Presentations 21.8 Percent Correct out of - SIR - 0 dB Impairment PB, 98 dB, flat, 101 APPENDIX B. CONSONANT CONFUSION MATRICES N - 8 2 66 16 - 1 4 2 2 - 23 36 - L y H WE - - - - 5 - - - - - 3 - - - - - - 4 - - 2 3 42 16 44 22 27 50 Three simulated normals P P T I B D G F TH S SH V II Z ZH CH J m N R W L y H WE 25 7 I 3 - T 7 50 8 - 5 2 - - - - - - - - 2 - - 38 72 - - - - - - - - - - 1 1 - - 6- -- -- -- -- -- -- - 2 4 - - - 3 - - 4 - - - 2 3 - - - - - - - - 1 - - 7 - - 5 - 2 1 - 2 25 20 - 2 3 4- - - 3 - 8 3 1 13 - 2 1 1 4 1 - - - - 21 9 65 11 20 1 - 1 2 18 8 648 - 5 2 11 3 1 - 1 8 4 1 25 - 37 14 22 33 37 1 B D G F TH S SI V IN Z ZI CH J N I R V L Y I 5 3 25 - 2 43 13 - 2 4 9 24 20 2 1 - 4 - - I 1 3 4 12 - 4 - - - - 1 1 - 7 - - 7 1 1 I 1 2 2 - 6 2 2 10 - - 10 36 - - - - 7 22 7 9 40 2 - 3 40 - 3 36 - 39 - - 33 - 20 2 - 22 12 2 3 101 34 46 40 5 - - 28 7 ----- 1- -- 52 -1- -- - - - 8 6 14 15 9 46 116 14 - 1 7 24 - 2 3 -- -- - -4 - ---- - 77 1 - 38 - 5 69 - -- 12 37 - 2 - 50 - 11 6 I 6 - 2 - I - - 5 1 9 4 17 - - 3 - - 1 4 - 43 - 1 --- -- - - -- 54 39 14 78 54 52 1 - - 1 - 1 5 - 3 I 1296 Presentations I 27 9 - - - - 5 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 -1 4 1 3 4 2 115 I - 65.9 Percent Correct out of -I - 1 2 1 6 -10 1 1 1 5 3 1 - - - 3 - 2 2 1 1 - -2 2 - - - 2 - - - 2 I - R W 2 - - 3 - 1 - - 1 - - - - - - - 2 42- WE - 22 H - - 10 - - T - - - 3 2 - L - - 17 1 6 V - - - R - 23 I 1 - 1 17 7 - 6 - 6 31 2 - 22 - 57 83 56 47 63 - 7 - 1 W - - - 1 - 2 - N 54 - - 2 - 3 1 - 3 4 3 - 14 - 2 1 - 2 3 2 - - - 1 1 J 54 54 54 54 - 12 11 17 - CH 54 1 - - - 1 - ZE - - - - Z 54 54 54 54 54 54 54 - - - - - II - 1 V - SH 648 Presentations of 12 - S - TH - F - 1 2 5 13 3 G - 3 6 1 - D - - 13 22 22 B - 3 - I - T - P T K B D G F TH S SH v II Z ZI CH P 35.3 Percent Correct out - Impairment PB, 98 dB, EFE, Quiet 102 1 54 54 54 54 54 54 54 54 54 54 11 54 25 1296 APPENDIX B. CONSONANT CONFUSION MATRICES - 1 71 35 1 10 - -18 - 2 4 1 1-- - 1 1 - --- - - 6 - - 1 - 40 16 2 1 17 9 31 - 35 22 - - 17 1 - - - 5 6 1 - 2 y N WE - 14 - - - - 13 - 7 4 - - - 5 - - - 2 - - - 3 - - - - 4 - 42 68 40 85 75 2 - S - - 1- - - - - - - - - 47 - 5 2 - -2 - - - - -2 - - 2 1 49 90 7 39 2 6 - - 1- - 2 - - 2 - SN V - 2- 17 IN 18 26 1 - 23 -- N R V L Y H 3 - - - - - - - - - - - - - - - - - - - 2 11 - - 3 - 4 - - 3 3 2 - 11 1 - 8 - 9 2 36 5 35 9 28 - 82 S - I 6 - - - - - - - - - - 1 - 2 - - - 1 - 4 13 - 3 10 - - 49 850 4 - 7 - - - - 1 3 17 2 5 - 52 - 3 - - -38 -15 -53 1 34 17 - - - - ---- - --- - - 11 75 - - -- - - --- - - - -- - 45 102 42 7 11 3 - 648 1296 Presentations N 2 11 - 9 J - - 40 CH - - 50 ZE - - 26 Z - - L S - - ZH CN M N R v - -37 - J 59 14 - - 6 - 26 - 9 27 - 9 10 4 - - z 4 3 I - - 1 - 11 29 - 1 1 8 - - 10 38 46 69 4 -1 - - - 1 8 5 - - 6 1 3 5 9 1 34 - 47 4 4 - 22 - 65 55 56 WE - - 34 3 -23- 1 - 4 4 - - 4 - 3 1 42 - 4 4 1 - - 3 2 - - 2 16 -- 1 24 - I 2 1 3 - 2 - I 1 - S - 1 - 14 1 TH - 1 - I 4 4 F - 5 - - 1 - G 6 - F TH S SH V IN 52 D - - 1 2 3 - - B D G 27 - B -7 54.7 Percent Correct out of - K I 3 - - - P T T 1 - - 1 Three simulated normals P -1 - - 29 - 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 27 - 28 1 1 H WH Y 1 - 25 L 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 5 64 64 64 64 S4 64 16 1296 - 11 V 8 - - 23 R - WE 1 I - - - 1 1 I 43 3 10 - - 1 - 4 - 2 7 16 - 2 2 1 4 - 1 1 2 K - - 7 17 2 6 CH - - 4 - 8 ZH - - Z - 2 IN - - 1 V - 5 3 3 11 3 1 2 9 12 - R V L y H 1 - - 1 2 1 - 1 2 SH - 1 S - - 2 7 4 1 1 TH - 16 3 F - 15 - z ZN CH J G 4 - - D - 3 B - P T K B D G F TH S SH V IN I T 648 Presentations 29.6 Percent Correct out of - P 0 dB SIR - HFE, Impairment PB, 98 dB, 103 APPENDIX B. CONSONANT CONFUSION MATRICES - - - - - - - - - - - 6 1 3 3 - - 1 - 2 - - - - - 2 - 1 5 1 2 2 - - 2 9 2 - 3 - 1 1 - 1 - - 18 12 -9 - - - - -- - 1 - 13-1 - 2 8 2 - 29 0 33 26 Two simulated normals N N R V 11 - - 5 2 - - - 23 7 - - 1 13 - - 1 1 - 13 1 6 - - - - - - - 2 2 - 5 - 2 - 5 1 1 1 - ZN CH J K 1 8 I - - 7 - - - 2 2 - - - - - -2 1 - - - 2- - - - - -2 - - - - -2 - - - 1 27 - 1 -1 - 33 8 1 3 - - - 2 1 11- 23 10 11 - - 27 43 17 39 - 11 8 6 - - - 3 2 1 13 36 47 17 3 - - - - - - - 2- - 24 10 2 23 - 4 1 27 17 10 34 43 3 9 30 - - - - 53 35 1 3 46 432 864 Presentations - Z 34 2 - 21 Ill 22 - IN - 1 1 2 V L y N WV 11 - - 7 - - SH 1 - - - - 2 - 11 2 - 1 I 4 2 - - 2 1 6 18 - - 15 - - - - 6 28 18 R V L Y N 2 5 1 - 5 1 2 - 2 2 - 2 2 3 7 - 1 5 - - 5 2 - 11 1 1 - - - - 2 - - 7 4 1 8 1 1 1 - 1 - V 1 3 2 - - 10 - - 53 39 36 36 36 36 36 36 36 36 36 36 36 36 36 1 36 36 36 36 36 36 36 - - 4 713 30 78 - 2 2 70 - - J 1 22 - - - F TH S SH V IN z ZH CH - 3 6 S - - 1 - - - 6 - TH - B D G - - F - 1 4 16 - - - - - G - I D - 2 29 B - 10 - I - P T 4 46.6 Percent Correct out of - T 5 1 - - P 1 - - 5 15 2 - 15 - 4 - 10 - 20 - - - - 31 1 - 15 1 - 16 - 2 - - 1 -- 2 1 - - L y H WE - - 1 - - 3 13 - 2 1 - W 6 3 - R - -1 1 N 12 3 - 3 1 - - - - 2 4 - - 1 - - - I 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - 1 5 1 WE - 2 - N - 1 Y - 6 4 -I -1 L - 1 2 - R - 4 7 - I - 3 N V J - - - 2 CH - 7 ZN - - - Z - - IN - 3 - 2 - 2 - V 432 Presentations - - - 5 2 11 - - 2 12 1 - 5 1 2 - 5 4 SH - 1 S - - TH - - F - G - z ZN CH D - 5 10 2 B 39.4 Percent Correct out of - 2 1 I - P T K B D G F TH S SH V IN T Quiet - P PB, 104 dB, flat, - Impairment 104 - - 2 36 36 36 36 48 34 14 864 3 APPENDIX B. CONSONANT CONFUSION MATRICES 4 4 1 - 1 3 1 N WE - -1 -1 1 1 2 2 11 4 1 - - 5 15 39 5 25 - 18 1 3 2 - - 1 - - - - - 1 - - - 1 - 17 4 -- -- - 4 - 1 3 - 1 1 - 6 31 27 73 TH S SE 2 6 6 - - - 24 CH - - 10 1 - 2 - 25 15 13 1 3 - 3 2 2 1 - 4 - - 6 11 9 - - 2 - 9 5 -4 3 - - - - - - - - 1 1 - - 1 1 1 - - - - 7 - - - 1 - - - 1 - - 2 4 - - 58 13 70 It 21 32 64 14 N N R V L Y N - - - 2 1 - - 4 2 WE - - - - 1 19 7 2 - 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 1 1 19 - 1 36 22 - - - - 1 - - 1 6 - 3 3 2 9 2 - - - 1 - 2 -- 4 - 1 30 1 2 - - 5 - 3 10 - - 2 2 - 8 7 - -- 5 - - - 36 - - 18 - - 1 - -- 35 - 22 - - 9 22 - 28 0 9 30 66 1-2 2 11 5 - - - I 4 2 - 66 45 - 5 18 432 864 Presentations J - 8 28 1 4 8 3 35 - 4 3 IN 6 - 6 - - - 10 - - ZN - - Z 6 2 9 1 - V - 2 F - - 1 G 2 - 17 6 - -- - 17 - - 4 13 2 - 20 17 2 2 -- 4 - - 3 3 10 - - - - - 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 - - 1 1 3 13 1 - 18 18 - 3 - 6 - 6 - D 4 3 - - - 14 - B 3 18 18 18 18 18 18 1 - - 12 2 - 2 K - T 1 - WE - L y - - N I R W 1 - J 1 42.9 Percent Correct out of P - R - 16 - - ZH CH 2 - - 17 - z 2 - 11 - F TH S SN V IN - - B D G - 3 1 - -1 Two simulated normals I - 12 Y 1 1 3 2 L - 1 - 6 - 6 17 1 35 2 - 33 72 45 46 - 1 - - 3 V 1 36 36 36 36 - W L Y N WE R 3 R P T -1 - -3- 1 10 - - N 4 5 1 - 7 -12 1 1 - I 12 - - - N - 1 2 - 1 J - 3 1 2 CH - 2 -13 IN ZE ZN CH ZH - - Z - 1 IN - 2 2 V - 1 - S SN - - p T K B D G F TH S SN TH - F - G 432 Presentations - D - B - I 29.6 Percent Correct out of - T SdB - P 3 - Impairment PB, 104 dB, flat, 0 105 45 18 864 APPENDIX B. CONSONANT CONFUSION MATRICES 2 -- - - - -21- 1 - 3 1 - - - - - - 2 3 2 - 1 -- -- - 2 - 17 - 2 - 1 -6 1 - 1 16 1 3 22 53 36 13 6 24 - - - - 3 4 26 20 - - - - - 5 36 - - - 3 - 9 - - - - 22 - -- - 12 - -- 14 1 - - 5 -- -- -1 IN Z - - - 1 - - - - ZE - 44 26 33 56 24 66 - 19 36 17 14 17 18 18 18 22 8 11 N WE R V L Y -2 - - - - - - - 9- 6 6 - - - - - 1- - - - - 36 36 36 36 36 36 36 36 36 36 36 -1 -- 13-4 2 3 36 -9 - -2 - -7 1 -- - - -- - - - -- 1 2- - -1 351 - 1 29- - -1 36 - 2 - 1- - -4 1 - -1 - 530 --- 3- 2 37 17 2 37 30 49 432 864 Presentations - 29- - - 6 - 5 -32 -27 - 6 5 - 1 4 - 2 4 25 2 - - - 36 - 1 N --30- -5 4 N 36 - -- 2 18 12 J - -- - CH 24 3 3 4 4 - - - 15 4 - 17 - - - 3 21 - V - - - 36 4 3 SH - 23 - - 3 4 18 18 - S - 41 2 18 2 1 9 - - TH -35 - 1 2 - - F 2 - 20 - - G - D - 3 - B 1 - 3 2 - I - - - 2 8 2 - 36 - 30 -- T 1- 2 10 70.6 Percent Correct out of P 2 1 - Two simulated normals 4 - - - - - 2 - - - 1 1 18 18 - 24 - 4 - 12 2 1 18 18 18 18 36 36 36 36 36 36 36 36 36 36 36 36 - 16 - 31 - - 18 - 3 18 - - - 16 - - 34 - - 2 18 18 - - - 1 - 7 - 1 - - L y H WE - 18 18 18 1 2 21 7 3 - 4 -- -I R - -2 N I P T I B D G F TH S SH V IN Z ZN CH J N N1 R W L y H WE -1 18 18 18 18 18 - 16 - - -I - - 1 - H WE - - - Y - - 2 2 L - 1- 9 - V - - - I - - R - - - 1 8- - - - I - - - - - - - N - -0 1 - - I 1 J - - CH - 1 - - -- - - - - ZN - IN 6 Z - I - SH 432 Presentations - S - TH 16 11 - F - 1 G - 1 D 41.4 Percent Correct out of - 1 B - 4 - K - P T K B D G F TH S SH V IN z ZH CH T Quiet - P HFE, - Impairment PB, 104 dB, 106 42 41 44 32 62 47 39 40 19 9 864 APPENDIX B. CONSONANT CONFUSION MATRICES SH - - 1 - - - - - - - 4 - - - CH - 7 2 - - 2 1 - 1 2 - 2 - 2 - - 1 - - - - 1 1 - - 13 - - 1 7 - - - 1 - - I - - 1 2 - - 1 - 14 -16 1 7 1 1 7 - 1 - 2 - 1 - - - 1 -- - 4 -- 3-4 - 4 -3 2 2- - 3 1 1 1 - H Y 6 W -I 18 1 3 2 18 18 18 4 1 18 3 1 1 1 - 1 - - - 2 - 1 3 - 1 - 2 -2 82 1 2 1 6 2 1 - - 3- 3 - 15 2 - -- - 1 - 4 - - - 1-1 - 4 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 14 432 - 1 - 1 18 26 28 31 16 1 3 39 D G F TH S - 3 - - - 9 - - - 36 1 20 - 8 - - - 1 -- 34- - 13-- -- -- - -- - - - - - J - 8 - 11 - 7 - - - - --- - -36 9 - 2 - 4 N I R W 4 1 - -2 - 1 -- 1 - 1 -36 9 -- - - -- 32 49 24 3 - - - 1 - - 39 26 31 73 20 37 - - - - - 27 864 Presentations J N I R V L Y N - - - - - - - - - - - 3 - - 2 - 2 - - 1 - - - 6 2 - - - - - 3 1 3 - 2 2 1 1 321 -- -- -- - - -1- 29- 361 - 6- - - - - - - - - 1 - 30 CH - - 27 ZN 1 36 Z 2 7 IN 1 L y N WE 19 - - 10 2 I - 12 - - WN - - 5 220---- - - 26 I 8 23- - - - 4 1 21 2 2 6 2 4 3 - 4 - - 30 4 23 2 24 - 1 36 - 22 - 62 57 38 38 30 1 36 25 4 40 30 58 33 47 27 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 36 3 36 36 8 864 - 1 5 8 14 - - 4 30 - - - - 1 231 - - 1 - 7 - - 4 - 4 - - B D G V - 3 - I - SH 1 - - - B 8 - - - I 21 10 66.1 Percent Correct out of T P T 20 4 - - - - - - - 1 - - - 1 10 - - 2 - - - - - - 3 13 - 2 - - 1 - Two simulated normals P I 3 -5 - 4 - - 17 2 - 14 - - 12 2 1 - 13 - 1 - 28 1 - - 2 4 - 2 2 - 2 - - 1 18 18 - L - 1 V - - - 3 - - -I - - - 1 1 R - 8 6 2 I - 6 1 2 - 10 N - - 16 - J - ZE - Z 2 - 5 F TH S SH V IN Z ZN CH IN - V - S - TH - F G - D 432 Presentations - B 38.7 Percent Correct out of - 3 I SIR - P T I B D G F TH S SH v IN Z ZN CH J N I R W L y H WE T 0 dB - P HFE, - Impairment PB, 104 dB, 107 4