f
41
Hearing Research, 18 (1985) 41-55
Elsevier
HRR00594
Some aspects of temporal coding for single-channel electrical stimulation of
the cochlea
Robert A. Dobie * and Norbert Dillier
ENT-Clinic, University Hospital, CH-809! Zurich, Switzerland
(Received 3 October 1984; accepted 27 March 1985)
Estimates of the useful frequency range for single-channel electrical stimulation of the cochlea range from 400 to 4000 Hz.
Psychophysical studies in single-channel implant patients are relevant not only to the practical problem of designing stimulation
strategies, but also to questions of temporal processing of pitch in the normal auditory nervous system. Patients with single-channel
extracochlear devices participated in several experiments involving stimuli differing in fine temporal structure. Stochastic pulse trains,
in which the probability of pulse delivery (p) for a given cycle was less than 1.0, were readily discriminated from ordinary pulse trains.
Frequency discrimination using stochastic pulse trains differing only in fine temporal structure (but identical average pulse rates) was
as good as with ordinary pulse trains or sinusoids for P 0.5, but deteriorated rapidly for P < 0.5.
Discrimination of triangular and trapezoidal waveforms from square waves was surprisingly good: rise-times (for 0 to maximum
current) as low as 0.08 ms could be discriminated. Conversely, detection of jitter in pulse trains was almost an order of magnitude
worse. The results show that frequency discrimination for single-channel electrical stimulation of the cochlea is based on
discrimination of inter-pulse periods, and that pulse rates which would be unnatural for acoustically-evoked VIIIth nerve activity — up
to 750 Hz — are more useful for coding mid-range frequencies than low-rate stochastic simulations of normal VIIIth nerve firing
patterns. The waveform discriminations reported would be obscured by low-pass filtering even at 2000 Hz, and probably depend on
changes in relative synchrony among an array of VIIIth nerve units with different thresholds. In general, these results support the use
of analog coding schemes with relatively large bandwidth.
cochlear implant, temporal coding, electrical stimulation
Introduction
It has been asserted [11,15] that frequencies up
to 3-4 kHz can be usefully presented to deaf
patients by electrical stimulation of the cochlea
through a single channel. If this is so, it must be
attributable to some type of temporal coding, as
opposed to a place code, since a single-channel
system obviously cannot exploit the normal spatial
array of cochlear neurons along a high-to-low
frequency (base-to-apex) dimension. The evidence
presented for such a phenomenon is of three general types:
1. Measures of patients' abilities to discriminate
sinusoidal stimuli of different frequencies have
• Present address: Department of Otolaryngology, RL-30, University of Washington, Seattle, WA 98195, U.S.A.
in some cases shown significant discrimination
up to 1-2 kHz [12,1].
2. At repetition frequencies up to 400 Hz, patients
can distinguish among square, triangle, and sine
waves, indicating an ability to utilize waveform
information [14].
3. Measures of Speech recognition abilities are in
some cases reduced when frequencies above 900
Hz are excluded by filtering [11].
The ability of individual auditory neurons to
phase-lock and thus preserve information regarding stimulus frequency in their interval histograms,
at least up to 2000 Uz and probably higher, is well
known [24,16]. As Evans [4] has pointed out,
frequency cues corresponding to the place' of
activity in the peripheral auditory system as well
as to the fine temporal structure of the discharge
patterns are both available to the central auditory
0378-5955/85/$03.30 © 1985 Elsevier Science Publishers B.V. (Biomedical Division)
42
nervous system (CANS). In the normal auditory
system the voice pitch can be extracted from the
combined temporal and place information of an
auditory-nerve fiber population for a wide range
of intensities and additive background noise conditions [22]. The limits of frequency discrimination
based on only temporal information have been
difficult to study psychoacoustically because of the
near impossibility of excluding spectral information in auditory stimulation of normal subjects [7].
In addition, although the temporal firing pattern
of a single neuron contains adequate information
to determine rather high stimulus frequencies, it is
not clear whether the CANS can extract this information from a single neuron or whether the temporal firing patterns of an array of neurons must
be compared in order to make such an analysis.
Theoretically, single-channel cochlear implant
patients should be nearly ideal experimental subjects in which to study these questions. Spectral
information can, in a sense, be ignored. At least, it
can be assumed that the spatial distribution of
cochlear nerve excitation is no longer dependent
upon the stimulus spectrum. However, many (perhaps most) of these patients have severe losses of
both sensory and neural structures, and there are
typically practical difficulties in getting these patients to perform arduous, boring psychophysical
tasks. Thus, frequency discrimination results from
single-channel cochlear implant patients must be
viewed as minimal estimates of the ability of the
normal CANS to utilize temporal information.
Under conditions of auditory stimulation, even
for low frequencies (e.g., 100 Hz), a given unit
usually does not fire on every cycle of the stimulus, but in a stochastic fashion; as a firnt-order
approximation, there is a fixed probability of firing (much less than 1.0) for each cycle. Thus, the
interval histogram will contain peaks at values of
nT (n = small integer, T = period of stimulus),
with the size of the peaks decreasing in an exponential fashion. For low frequencies, whose
periods are greater than the neuron's refractory
period, at least some of the interspike intervals will
be equal to T. For higher frequencies (above about
1 kHz), only integer multiples of the period (2T,
3T, ) will be represented in the units interval
histogram.
If the CANS examines the temporal firing pat-
terns of an array of units, the composite interval
histogram will contain values as small as T, even
for frequencies above 1 kHz, where no single unit
could fire at the stimulus frequency. This is, of
course, the volley' theory of Wever [31].
There is very little information available regarding single unit firing patterns for electrical
stimulation, but it is known that the degree of
phase-locking is more precise (less jitter) than for
acoustic stimuli, and that maximum discharge rates
are extremely high -- up to 900/s ([17,8]; Van den
Honert, 1983, personal communication).
The relatively unnatural nature of the auditory
nerve firing patterns in response to electrical
stimulation has also been noted by Sachs et al.
[25], who have suggested the use of coding strategies incorporating lower pulse rates and stochastic
properties. Presumably, the central nervous system
(CANS) may find it difficult to make use of the
abnormally high rates of firing elicited by even low
frequencies (e.g., 500 Hz). Interspike interval
histograms of single units excited with low-probability electrical pulse trains (intervals T, 2T,
3T, ) are assumed to be similar to interval histograms of single units stimulated with periodic
acoustic signals of frequency 1/T. The difference
for the CANS however is that for acoustically
stimulated fiber groups originating from the same
cochlear region the temporal and spatial averages
are equivalent (ergodicity principle) whereas for
electrically stimulated fiber groups they are not. If
an instantaneous representation in a spatially distributed fiber group is a prerequisite for adequate
central pitch processing, then single channel stochastic stimulation probably could not efficiently
be used for conveying higher frequency information. If, on the other hand, some temporally averaging central analysis mechanism for spike intervals is available, then stochastic stimuli could be
used to provide the CANS with a more natural
input.
Frequency (pulse rate) discrimination between
appropriately chosen stochastic pulse stimuli can
be assumed to be based exclusively on differences
in fine temporal structure, since frequency and
probability of pulse delivery can be co-varied so
that resulting stimuli have identical mean interpulse intervals (or 'pulse densities'). Thus, stochastic pulse stimuli offer a unique test of the ability of
43
the CANS to use the kind of phase-locked temporal information which has Jong been known to
be present in single-unit firing patterns [24].
It is useful also to consider how a patient with a
single-channel cochlear or extra-cochlear implant
could distinguish among different waveforms
(square vs. triangle vs. sine). Clearly, spectrum per
se plays no role. If one waveform were more
effective in stimulating a larger array of neurons,
this could be the basis for discrimination, but if
the stimuli are first adjusted for equal loudness,
this would be expected to eliminate cues based
only on the size of the neural array stimulated.
The most likely way for such a discrimination
to be made is by the differing temporal distribution of spikes for the different stimuli. A square
wave will presumably fire all units whose
thresholds are exceeded, nearly simultaneously,
and the degree of phase-locking should be nearly
perfect, i.e., a given neuron's firing will be at
intervals which are precise multiples of the stimulus period. Conversely, the triangle wave, at the
other extreme, will fire low-threshold units slightly
earlier in every cycle than high-threshold units. In
other words, the degree of simultaneity of the
array is reduced. Similarly, each individual unit's
firing pattern, over a large number of cycles, might
show reduced precision of phase-locking, i.e. increased jitter'. This would be apparent from an
examination of an interval histogram, and could
also be the basis for discrimination of the different
waveforms. Note that the distinction could be
made based on a single unit's firing pattern only if
a sufficient number of cycles is presented to permit some kind of analysis of interspike intervals,
while an analysis of the degree of synchrony among
an array of units could theoretically be performed
after even a single cycle of the stimulating waveform.
Thus, studies of discrimination of fine temporal
structures of electrical stimuli are of interest for
two major reasons. First, they may assist in developing coding strategies for patients with cochlear
(or extra-cochlear) implants. Second, they offer a
unique insight into the ability of the CANS to use
temporal cues in the complete absence of
spectral-place cues.
Materials and Methods
Each of the patients who participated in these
studies had received a single-channel extracochlear
implant, with a single-ball electrode at the round
window membrane. Technical specifications and
basic psychophysical performance have been described previously [28,2]. Surgery has in each case
been performed more than one year prior to these
studies; during this time, thresholds for electrical
stimulation had been stable and patients were
using wearable encoding and stimulation units on
a regular basis (see Table I for summary of technical specifications). Neither of them is a `star patient'; no open-speech understanding has been
achieved without lip-reading.
Two patients provided most of the data to be
reported. One (U.T.) had suffered bilateral profound deafness from sudden hearing loss superimposed upon congenital sensorineural hearing loss,
at age 39. The other (E.P.) was prelingually deaf
secondary to meningitis in the first year of life.
Neither had any response to audiometric tones up
to 110 dB. Each of them participated in 5-10
sessions of about 4 h each. Some of the experiments were repeated on different days; in such
cases, the best performance will be reported and
noted as such. Although neither test-retest variability nor leaming effects were formally investigated, both appeared to be important in these
difficult and tedious psychophysical tasks. However, the results can be taken as at least minimal
estimates of performance, i.e., some implant patients can perform at least as well as reported
herein on the tasks to be described.
Each of the tasks to be described was also
performed by 3-6 normal-hearing subjects in a
sound-field setting. The stimuli were transduced
TABLE I
TECHNICAL SPECIFICATIONS
Electrodes:
90% Pt, 10% Ir, 1 mm ball diameter
Round window niche vs. M. temporalis
Stimuli:
Capacitively coupled pulses, 0.1-0.2 ms
duration. Pulse amplitude and timing detertnined from amplified and compressed
microphone signal
Signal transmission: Amplitude modulation, 12 MHz
44
by electrostatic headphones with a flat frequency
response from 20 to 20000 Hz. Pulses used in both
the acoustic and electrical stimulation experiments
were 0.1 ms in duration. Amplitude levels for
pulses were set for each experiment by a
determination of most comfortable level (average
of three trials) using a slide-potentiometer.
All stimuli were generated through a digital to
analog converter (12-bit resolution, AA11-K, DEC)
using a dual programmable 1 MHz Clock (KW11K, DEC). The algorithms were implemented in
Fortran and Macro-Assembler on a PDP-11/34A
(DEC). For experiments with normal hearing subjects a lowpass filter (5 kHz, 24 dB/oct, KrohnHite) was switched between the D/A-Converter
and the output attenuator. The maximal frequency
resolution for pulse output with a 1 MHz clock
obviously depends on the absolute frequency. 1000
clock cycles per pulse interval will produce exactly
1000 pulses per second (pps), 999 cycles 1001.001
pps, 1001 cycles 999.001 pps. The 1 pps-resolution
at 1000 pps is not sufficient for psychoacoustic
experiments with normal listeners as their
frequency discrimination lies in the same range.
For implant patients however this value is certainly sufficient. At lower pulse rates the resolution gets progressively better. For example, a pulse
rate of 100 pps is generated using 10000 cycles,
100.01 pps using 9999 cycles.
Four types of temporal discrimination were
studied. In each case, variation of some stimulus
parameter gave rise to a clear variation along a
particular perceptual dimension (high-low, sharpdull, or smooth-rough), so that stimuli could be
labelled and discriminated. Difference limens
(DLs) were determined using an adaptive two-alternative forced choice paradigm, in which a transformed up-down method [19] with seven reversals
estimated the point at which the probability of a
correct response was 0.707.
For a frequency discrimination task, for example, each stimulus pair contained a reference
stimulus and a higher-frequency test stimulus, in
random order; the subject's task was to identify
the stimulus pair as either high-low or low-high.
All stimuli were 800 ms pulse trains or waveforms
with 25 ms rise—fall times, used to modulate the 12
MHz carrier for transcutaneous transmission. The
paired stimuli were separated by 400 ms; after
each stimulus pair, a delay of 3-5 s (depending on
the subject's response time) is imposed. If no
response is made after 5 s, the same stimulus pair
is repeated.
Stochastic pulse stimuli (probability and frequency
discrimination)
An ordinary pulse train at f = 500 Hz may be
considered to have a probability of 1, i.e., a pulse
is generated every 2 ms. If the probability is reduced to 0.8, there is an 80% chance of a pulse
being delivered at each 2 ms interval. In one
second, 400 pulses would be delivered (500 x 0.8),
and an interpulse interval histogram would reveal
intervals of 2, 4, 6... ms. Such a stimulus (500 Hz
at P = 0.8) was always heard as rough compared
to an ordinary pulse train (P = 1.0). Individual
800 ms stimuli were generated immediately prior
to delivery of each stimulus pair, so that, for
example, each presentation of 500 Hz at P= 0.8
was unique.
In one set of experiments, a probability DL'
was determined. The reference stimulus was a
pulse train at P= 1.0 (smooth) and the test stimuli
were stochastic pulse trains with the same base
frequency but stepwise decreasing probabilities of
pulse generation (rough).
Although stochastic pulse trains were heard as
rough, this did not prevent an appropriate labelling along a high—low continuum. Frequency discrimination experiments were performed in which
a reference stimulus (e.g., 500 Hz at P= 0.8) was
compared to other stochastic pulse stimuli with the
same mean pulse rate and interpulse interval (in
this case, 400/s and 2.5 ms). The test stimuli
would have higher base frequencies but lower
probabilities of pulse delivery (e.g., 525 Hz at
P= 0.762; 550 Hz at P= 0.727). Typical stimuli
are shown in Fig. 1.
Gap detection
When subjects discriminated ordinary from stochastic pulse trains, they were, in one sense, detecting gaps in an otherwise continuous pulse train.
Thus, for comparison, a simple gap detection task
was included. The reference stimulus was an
ordinary pulse train 800 ms long; the test stimuli
had gaps inserted at the midpoint of the stimuli.
One subject (E.P.) also performed a gap detection
45
A
STOCHASTIC PULSE TRA1HS
E00 Hz/p=0.8
667 Hz/p.0.6
1
e.
Time (msec
1.0
200
B
698. Hz/p.8.8
324
8.6_
68
9
0.8
8
ie '
Time 1msec)
1.8
667. Hz/p.9.6
8.5
239
199
39
0.9
20
5
Time (»sec)
5
ie
Fig. 1. Ordinary pulse trains at 500 and 667 Hz differ both in
inter-pulse interval (2 vs. 1.5 ms) and in pulse density (500 vs.
667 pulses/s). (A) Reduced probability of pulse generation
(P = 0.8 for '500 Hz' vs. P = 0.6 for '667 Hz') yields a stimulus
pair differing only in fine temporal structure; all interpulse
intervals are, for the '500 Hz' stimulus, multiples of 2 ins, and
for the '667 Hz' stimulus, multiples of 1.5 ins. The pulse
density is the same for both stimuli: 400 pulses/s. (B) The
interpulse interval histograms for the stimulus pair shown in A.
Numbers above the histogram bars denote number of corresponding intervals within a 1 s stimulus. The total count for the
'500 Hz' stimulus is 399, for the *667 Hz' stimulus 404.
task in which narrow-band noise at 250, 500 or
1000 Hz (316 Hz bandwidth) was the stimulus
presented with or without a variable gap.
Waveform discrimination
Discrimination among sine, square, and triangular waves in single-channel electrical stimulation
must depend an temporal differences in the waveform rather than spectral differences per se. Thus,
we considered the square and triangular waves to
be extremes of abrupt and gradual current change,
respectively, and constructed intermediate trapezoidal waveforms to test the Limits of discrimination of rate of change of current (see Fig. 2).
Since threshold for electrical stimulation may in
some cases be more closely related to charge delivered than to current, it is useful to consider the
functions relating total charge to time for different
stimulus waveforms. Ignoring the effects of charge
leakage for the sake of illustration, the instantaneous charge is proportional to the integral of the
current waveform. Thus, a square wave (current)
yields a triangular wave (charge) and a triangular
wave (current) yields a parabolic wave (charge).
Fig. 3 illustrates this point: the temporal differences between the charge waveforms are much
less than for the corresponding current waveforms.
The exact values of current and voltage across the
electrode are not known as the electrode impedance cannot be measured directly. However in
vitro measurements of the electrode characteristics
indicate mostly resistive behaviour in the frequency
range of interest and therefore nearly equivalent
waveforms for current and voltage. In this context
it might be important to note that the stimuli are
AC-coupled via a capacitance of 180 nF which
acts as a high pass filier emphasizing rapid waveform changes. If it is true that the effective stimulus is the integrated current per period rather than
the instantaneous amplitude then this integration
would counterbalance the differentiating effect of
the coupling capacitor.
It was recognized that loudness must be controlled in an experiment of this type; at equal peak
current levels, a square wave delivers twice as
much charge per half-cycle as a triangular wave.
For each experiment, MCL was obtained for each
of 10 stimuli to be used (ranging from square to
triangular in equal steps). Thus, each stimulus was
46
STIMULUS CURRENT ANO CMRAGE WRVEFORM5
RT/PERIOD
INTEGRATED CURRENT (CHARGE) HAVEFORMS
0.250
0.225
0.200
0.175
0.150
0.125
0.100
0.075
0.050
0.025
0.000
e.o
1.0
TIME 1051
0.2
0.4
TIME IMS I
0.6
Fig. 3. Charge waveforms for rectangular and triangular current waveforms.
2.0
Fig. 2. Variation of rise-time/period from 0.0 (rectangular
waveform) to 0.25 (triangular waveform). The curves are scaled
in amplitude to yield equal areas. Thick lines denote stimulus
(current) waveforms, thin lins are the integrated curves corresponding to charge over time.
presented at MCL. The current levels required to
equalize loudness were surprisingly close to those
predicted for an equal-charge model (Fig. 4).
In each waveform experiment, the reference
stimulus was a square wave, and the test stimuli
were trapezoidal; with gradually increasing rise
time/period, until at the extreme (rise time/
period = 0.25) a triangular wave was produced.
of that neuron's threshold more rapidly than a
triangular wave, and would thus be expected to
elicit a Spike with more precision on each cycle,
i.e., better phase-locking.
This condition can be simulated using a pulse
train with imposed jitter (Fig. 5). The reference
stimulus for these experiments was an ordinary
pulse train (e.g., 500 Hz) in which all the interpulse intervals were exactly 1/frequency. The test
stimuli were pulse trains with gaussian distributed
interpulse intervals with the same mean as the test
stimulus and increasing standard deviation (S.D.)
(e.g., 2 ms t i • 0.05 ms S.D., i = 0 ... 9).
Restilts
Jitter discrimination
Discrimination of waveforms (e.g., square vs.
triangular) could be based on changes in the degree of synchrony in a single neuron's interval
histogram (temporal averaging) and/or on changes
in the synchrony of an array of neurons (spatial
averaging). Assume that a given neuron's threshold
of response is not absolute, but has some variability. A square wave would pass through the region
Stochastic pulse stimuli
Temporal discrimination results are summarized in Table II. Normal listeners very readily
detect slight decreases in probability of pulse delivery. With no training, DLs of 0.05 are typical;
in other words, P = 0.95 is distinguishable from
P =1.0, at frequencies from 80-1000 Hz. With
electrical stimulation, U.T. obtained DLs from
47
EOURL-LOUONESS LEVELS FOR DIFFERENT WAVEFORMS tU.T.I
500 Hz PULSE TRRIHS WITH JITTER
sd (es«)
11111111111111111111111111111111111111111111111111
1111111111111111111111111 1111111111111111111111111I
1111111111111111111111111 111111111111111 111 11111 111 1
11111111111111 1111111111 II 11111111111 111 11111 111 II
11111111111111111III II II I 1111111111 111 11 I II 111 11
0.0
Time (msecl
0.
0
0.25
0.5
0.75
1.0
100.
Fig. 5. Pulse trains (500 Hz) with varying amount of jitter. The
pulse intervals are gaussian distributed around a mean value of
2 ms with standard deviations (S.D.) of 0 (regular pulse train),
0.25, 0.5, 0.75 and 1.0 ms (from top to bottom). Only 100 ms of
the 800 ms trains are displayed.
O
Mo
13.1-080.
4-1.250.
›FH0(500.
(F* 1000.
0.05
0.10
0.15
BISE TIME/PERIOD
0.20
0.03 to 0.11; E.P. from 0.12 to 0.30. U.T.'s DLs
were all less than 0.05 from 125 to 500 Hz. Both
showed slightly worse performance at higher frequencies. At 750 Hz, their DLs were 0.11 (U.T.)
STOCMASTIC PULSE RATE DIFFERENCE LIMENSI 3 NORMALS
0.25
8
P.
EQUAL-LOUDNESS LEVELS FOR DIFFERENT WAVEFORMS (E.P.I
a-0.2
X--S 0.
8
+- 0.8
2-Z0.8
X-1(1.0
2
B
O
w2
cc
70
0.0
8
•b.o
250.0
550.0
FREGUENCY (HZ)
750.0
1000.0
FRED (HZ)
Ch4D80.
A-4125.
01-0125.
£-+250.
+-+500.
»4(750.
411-• t 000.
8
>H0(500.
40-6750.
4-+1000.
A
(c
ue
2..
-I 8
111.05
CAD
0.15
RISS TIME/PERU:1D
8
L0
Fig. 4. Normalized peak amplitude levels for equal Ioudness
sensation versus rise time/period at different frequencies. (A)
Data for U.T.; (B) E.P.
FIRING PR08R8ILITY
tlo
Fig. 6. Relative DLs (df/f, in percent) for stochastic pulses:
mean values for 3 normal subjects.
48
TABLE II
DIFFERENCE LIMENS - TEMPORAL DISCRIMINATION
Freq.
(Hz)
Stochastic pulses
Gap DL
Pulse
prob.
Jitter DL
(ms) (S.D./T, %)
1.0
0.8
0.6
0.4
0.2
0.6
0.6
0.3
0.4
0.5
0.3
2.2
1.5
0.4
0.5
0.5
0.1
1.0
0.7
0.5
0.5
0.5
0.5
0.01
0.01
0.01
0.02
0.03
0.02
0.18
0.08
0.03
0.02
0.01
0.01
4.5
3.3
1.5
1.0
2.7
2.0
0.07
0.12
0.02
0.03
0.14
0.17
2.20 (27.5)
1.25 (15.6)
0.77 (19.2)
0.27 (13.5)
0.40 (30.0)
0.05 (5.0)
0.11
0.14
0.17
0.01
0.15
1.53 (19.2)
1.30 (32.5)
0.37 (18.3)
Normal subjects (mean values)
80
125
250
500
750
1 000
0.01
0.01
0.01
0.01
0.01
0.01
0.4
0.3
0.4
0.2
0.3
0.2
0.3
0.3
0.2
0.1
0.2
0.1
0.4
0.3
0.4
0.2
0.2
0.2
U.T.
80
125
250
500
750
1 000
0.11
0.04
0.06
0.03
0.11
0.07
2.7
7.5
3.8
7.0
7.0
25.0
32.0
5.0
5.2
8.0
24.0
28.0
11.3
6.3
5.0
20.0
E.P.
80
125
250
500
750
1 000
0.12
0.18
0.22
0.30
0.30
10.8
17.5
23.3
22.5
31.7
61.7
25.0
29.2
23.3
30.8
33.3
and 0.30 (E.P.). Both patients reported the stochastic pulse trains to be rough, and could clearly
distinguish the rough-smooth dimension elicited
from either loudness or pitch.
Pulse rate discrimination for stochastic pulse
stimuli is, for normal listeners, essentially as good
as for ordinary pulse trains, even for probability as
low as 0.4. With about 1 h training, normal subjects achieved relative DLs (expressed as df/f, in
percent) of less ihan 1% for all frequencies and all
probabilities > 0.2 (see Fig. 6, for example).
This is better than would be expected for
sinusoids (especially at low frequencies, where discrimination of, say, 125 from 126 Hz would be
exceptional) and is probably attributable to the
rich harmonic structure of the stimuli. A dick
train at 125 Hz contains harmonics at 2 and 4 kHz
which can be discriminated from harmonics at
2.015 and 4.032 kHz in a 126 Hz dick train. Even
at very low probabilities of pulse delivery, this
harmonic structure persists, so that this task can-
25.3
10.0
30.0
23.3
Noise
(ms)
Waveform
(RT/T)
Pulses
(ms)
Rate discr. (df/f, %)
16.7
5.0
4.0
60.0
1.8
55.0
(1.47)
(1.04)
(0.83)
(0.83)
(1.0)
(1.0)
0.14 (14.2)
not be considered a pure test of temporal discrimination in normal subjects.
As expected, the electrically stimulated patients
did muck worse. For ordinary pulse trains, U.T.
typically had relative DLs (df/f) of 3-7% up to
750 Hz, about the same as had been obtained with
sinusoidal stimuli. With stochastic pulse trains (P
= 0.8 or 0.6), she continued to hear the test stimuli
as higher in pitch than the reference stimulus, and
performed nearly as well as she had with ordinary
pulse trains. When probability was reduced further, performance was severely degraded (Fig. 7).
E.P.'s performance was worse, but showed similar pattems (Fig. 8): up to 500 Hz, he performed
nearly as well for P = 0.8 as he had for ordinary
pulse trains (or for sinusoidal stimuli). At 80 and
125 Hz he did as well with P = 0.4 and even
P = 0.2.
Gap detection
U.T. was able to detect gaps as small as 1.0 ms
49
STOCHASTIC PULSE RATE DIFFERENCE LIMENS, U.T.
STOCHASTIC PULSE RATE DIFFERENCE LIMENS: E.P.
P.
P-
+ 0.2
*-x0.4
2-Z0.8
)1(-* 1 . 0
0.0
`b.o
500.0
FREOUENCT (HZ)
250.0
5.00.0
FREOUENC (ND
750.0
10e0.0
8
FREI] (HZ)
FREI] (HZ)
0-1080.
0-4125.
0-4080 .
A-4 I 25.
›F-X500.
4-0750.
>H<500.
0.-0750.
4.-1'1000.
+-* 1000.
os
=
8
0.2
0.0
0.8
FIRING PROBRBILITT
0.0
1.0
0L0
0.2
0.0
dAN
FIRING PROEIRRIL1TT
de
1.0
Fig. 7. U.T.'s relative DLs (df/f, in percent) for stochastic
pulses.
Fig. 8. E.P.'s relative DLs (df/f, in percent) for stochastic
pulses.
(in a 500 Hz pulse train); her DLs for 250 Hz and
1000 Hz were 1.5 and 2.0 ms, respectively. E.P. did
less well, with gap DLs of 4.0 ms (500 Hz), 2.7 ms
(750 Hz), and 1.8 ms (1000 Hz). Performance for
detection of gaps in narrow-band noise (E.P. only)
was much worse, even after some practice: 60 ms
at 500 Hz, 55 ms at 1000 Hz.
and vice versa. Interestingly, on a second day of
testing, she reversed this labelling: square waves
were reported as sharp and triangular waves as
dull. Her ability to discriminate stimuli consistently along this dimension was unimpaired.
Perhaps sharp-dull was not an apt choice to describe the perceptual continuum heard by the patient, but it was (on each occasion) her choice.
DLs for waveform variation ranged from 0.02
(rise-time/period) for 250 Hz to 0.17 at 1000 Hz.
Her equal-loudness levels approximated the expected equal area function except for 1000 Hz
(Fig. 4A).
E.P. performed nearly as well as Ü.T. although
he was not able to label the differences in perception other than "sounds different". He chose to
perform the test at a level above his usually set
Waveform discrimination
U.T. was able to distinguish square wave from
triangular wave stimuli, even up to 1000 Hz. Prior
to formal DL measurements with loudness matching, she was able to label these as sharp (triangular) and dull (square). This distinction was subjectively different from loudness, and persisted even
when one stimulus was presented near threshold
and the other near uncomfortable loudness level,
50
most comfortable level. Again the loudness equalization was performed for every new frequency and
test condition. The equalized amplitude deviated
systematically from the expected area-function
(Fig. 4B). There was one exceptionally good performance at 750 Hz with 100% correct responses
for all waveform differences greater than the
minimum (rise-time/period = 0.025). At all other
tested frequencies the DLs were between 0.11 and
0.17 (see Table II).
Jitter discrimination
Both U.T. and E.P. heard stimuli with imposed
jitter as rough compared to the smooth sound of
ordinary pulse trains. This was different from the
sharp-dull dimension elicited by waveform variation. U.T. performed slightly better, with jitter
DLs (values in parentheses denote the jitter ratio
in percent: pulse interval S.D./mean interpulse
interval) of 2.2 ms (27.5%) at 80 Hz, 1.25 ms
(15.6%) at 125 Hz, 0.77 ms (19.2%) at 250 Hz, 0.27
ms (13.5%) at 500 Hz, 0.40 ms (30.0%) at 750 Hz
and 0.05 ms (5.0%) at 1000 Hz. The outcome at
1500 Hz was 0.04 ms (6.0%) (not shown an Table
II).
E.P. commented the jittered stimuli as "the
computer is crazy". His DLs were 1.53 ms (19.2%)
at 125 Hz, 1.3 ms (32.5%) at 250 Hz, 0.37 ms
(18.3%) at 500 Hz and 0.14 ms (14.2%) at 1000 Hz.
There were no measurements at 80 and 750 Hz.
Discussion
The ability of implanted patients to distinguish
stochastic pulse trains from ordinary pulse trains
could be considered a form of gap detection. When
P = 0.95 at 500 Hz, for example, most of the
interpulse intervals are 2 ms, but in a pulse train
800 ms long, the probability of at least one interval 6 ms (3 times the period) is about 0.61. The
probability of an interval of 8 ms or longer in such
a stimulus is only 0.05.
Thus, discrimination of such stimuli from an
ordinary pulse train, as was achieved by one of our
patients (U.T.) would be equivalent to detecting a
`gap' of about 4 ms (since the normal period of the
pulse train is 2 ms, an interpulse interval of 6 ms is
equivalent to an additional gap of 4 ms).
Even E.P., who in this as in most other tests
performed worse than U.T., was able to discriminate stochastic pulse trains of 750 and 1000
Hz, with P= 0.70, from ordinary pulse trains. The
maximum intervals present in such stochastic
stimuli would be 7.-8 ms in most cases, yielding
maximum gaps of 6-7 ms.
This is much better than the 22-157 ms (mean
= 52 ms) reported by Hochmair-Desoyer et al.
[13] for gap detection using broadband (250-1000
Hz) noise bursts in 12 implant patients. However,
our patients' performance in gap detection for
pulse trains was also much better than this. One
patient (E.P.) also performed gap detection for
narrow-band noise, and his performance was comparable to that reported by Hochmair-Desoyer et
al. [13]. It seems, therefore, that gaps in very
regular stimuli (pulse trains) are much more readily appreciated than gaps in irregular stimuli (noise
bands).
Figs. 7 and 8 show the pulse rate discrimination
performance of U.T. and E.P. for ordinary and
stochastic pulse trains. For ordinary (P = 1.0) pulse
trains, their performance is the same as for
sinusoidal stimuli (see also [29]). For slight reductions in probability of pulse delivery (P = 0.8 or
0.6), performance was unaffected. Recall that in
this task, subjects had to correctly identify as high
or low stimuli with identical average pulse rate,
differing only in the distribution of interpulse intervals (e.g. 2, 4, 6, ... ms for base frequency = 500
Hz; 1.82, 3.63, 5.45, ms for base frequency = 550
Hz).
Not only do these data unequivocally prove
that the CANS is able to utilize such temporal
pattems for frequency discrimination; they also
remove any doubt about confounding of pitch and
loudness, since the test stimuli all contained equal
numbers of identical pulses.
Performance deteriorated rapidly for P smaller
than 0.5, suggesting that these patients were primarily relying upon the periods of the base frequencies themselves in making frequency discriminations, rather than the higher multiples of these
periods which were increasingly represented for
low-probability stimuli. Thus, a relative DL (df/f)
of 5% at 500 Hz represents a discrimination of 2
ms vs. 1.9 ms, or an absolute DL of 0.1 ms. The
most 'realistic' of our stochastic pulse stimuli, in
terms of imitation of VIIIth nerve spike trains,
51
were probably those with low probability of pulse
delivery, yielding average rates under 200/s, e.g.,
P = 0.4 at 500 Hz, P = 0.2 at 1000 Hz. These were
poorly discriminated, and sounded very unpleasant to our patients.
We have to conclude from these results that
single channel stochastic pulse stimulation (imitating the temporal firing distribution of single auditory nerve fibers) does not provide a truly `naturaT pulse code and that probably independent and
different versions of stochastic pulse trains would
have to be supplied to different neurons to mimic
also the spatial discharge pattern distribution. This
would however require multiple stimulation channels for each frequency band to be transmitted,
and is probably unrealistic. From the single-unit
data available, it seems likely that electric pulse
trains up to at least 600-900 Hz elicit cycle-forcycle firing [9,30]. While this is indeed unnatural
vis-ä-vis acoustic stimulation, the CANS is able to
use this information to make frequency discriminations based on fine temporal structure.
Neural models capable of such an analysis of
incoming VIIIth nerve data have been described,
and postulate an array of units tuned to `characteristic periods'. These units could receive separate straight-through and delayed versions of the
VIIIth nerve spike train (Licklider's neural autocorrelator, [20]), or could be oscillators with positive feedback loops of varying temporal path length
[18]. In either case, such neurons would respond
maximally when the interval between incoming
spikes (T) is the same as the pathlength of the
feedback loop, or the delay line. Even if no single
VIIIth nerve neuron fires at rates of 1/T, a group
of neurons from the same cochlear locus could
converge upon a single second-order neuron, so
that aggregate input to the latter cell would be a
spike train at f = 1/T. Godfrey et al. [5] have
described cochlear nucleus units which fire cyclefor-cycle with dick trains up to 800/s. Such units
presumably receive convergent input from several
auditory neurons and could provide regular spike
train input to higher order neurons actually performing period analysis. To our knowledge, period
analyzing units like this have not been described in
die brainstem, but could be sought with electrical
stimuli like those used in this study. Similar units,
binaurally innervated with "characteristic interau-
ral delays" have been described in the superior
olivary complex [6].
Models like this would detect only a particular
period T, and would not be capable of analyzing a
single stochastic spike train with interspike intervals of nT unless one postulated multiple delay
lins (or feedback loops) for each unit with delays
of T, 2T, 3T,... etc. This seems unnecessary and
certainly less parsimonious than the models described above. In addition, our patients' poor performance for stochastic pulse trains with P < 0.4
suggests that they relied mainly on the primary
intervals present in the stimuli and were not able
to use the period multiplier (2T, 3T, 4T, ) very
well.
It is often argued that frequency discrimination
based on temporal features of electrical stimulation (` rate pitch') is limited to the region below
4-500 Hz [27,26]. Indeed, pitch scaling clearly is
in many cases random above this level [10]. However, frequency discrimination as good as 10%
(df/f) has been reported for frequencies as high as
1000 Hz [28,12] or 2000 Hz [1]. Muller [23] argues
that loss of pitch-labelling or scaling ability (above
500 Hz) does not necessarily mean that information from higher spectral regions is wasted in
electrical stimulation of the cochlea; cases where
speech discrimination is impaired by low-pass
filtering at 900 Hz [11] support this point of view.
Another argument for the inciusion of relatively
high frequencies in coding schemes is the ability of
patients to discriminate waveforms which differ
(in the frequency domain) only at harmonics of
the frequency used. One of our patients (U.T.)
could clearly distinguish square from triangular
waves independent of intensity manipulations, up
to 1000 Hz.
Two possibilities for neural encoding of such
waveform differences were briefly mentioned in
the introduction. Most likely is that the degree of
synchrony among a population of VIIIth nerve
units (greater for square waves, less for triangular
waves) provides the cue for discrimination. Fig. 9
illustrates this idea. If one assumes that units A, B
and C have different thresholds, it can be seen that
they will discharge together for the square wave,
but at different phases of the triangular wave
stimulus. The neural substrate for this phenomenon has been provided by the finding of Van den
52
Honert and Stypulkowski [30] that individual
VIIIth nerve unit thresholds, independent of
cochlear place, have a range of about 12 dB. Loeb
et al. [21] have similarly interpreted perceptual
differences of electrical stimulation with
equal-charge pulses of different durations in terms
of a spatial cross-correlation mechanism involving
units with different Eiring thresholds.
An alternative possibility would be that a single
neuron had not a precisely determined threshold,
but instead a range of possible thresholds (mean
some standard deviation). The CANS could then,
by examining the interval histogram of such units,
distinguish between various degrees of synchrony.
For illustration, Fig. 9 can again serve, if A, B
and C are now thought of as discharges (at different times) of the same unit, whose range of possible thresholds is shown.
This latter hypothesis seems less likely, since no
one has reported individual unit threshold variability for electrical stimulation. In fact, the extremely steep rate-level curves reported by Kiang
and Moxon [17], in which the dynamic range
(from spontaneous rate to maximum) was only
about 4 dB, argue against this. An indirect test of
this hypothesis was provided by the experiments
involving jitter DL: it was assumed that pulses at,
say, 250 Hz with imposed jitter would simulate the
response of neurons with variable thresholds to a
gradual waveform (triangular). In fact, performance on the jitter DL was much worse than
would be required to make the required waveform
discrimination. U.T. could distinguish an ordinary
250 Hz pulse train from one with a jitter ratio of
19.2%, i.e. 95% of interpulse intervals were between 2.46 and 5.54 ms (mean ± 2 S.D.). Waveform discrimination at 250 Hz, on the other hand,
was possible with a DL (rise time/period) of only
0.02; a trapezoid with a rise time of 0.08 ms could
be distinguished from a square wave (see Fig. 10).
At 500 Hz, the results were similar: the waveNAVEFORM AND JITTER DL. U.T. 250
HZ
0
cc
o.a
1111111
111111)
1114 ,
IM
uadi
..didil i111111111111111...:
PULSE GENERATION WITH DIFFERENT NAVEFORMS 1500 HZ)
dS
5
c
8
o.o
0.0
TIME INS)
2.0
TIME IMS1
0.0
Fig. 9. Hypothetical firing patterns for three units with different thresholds for rectangular (left) and triangular (right) stimuli
(500 Hz). (A) Low threshold-unit. (B) Medium threshold-unit.
(C) High threshold-unit.
i'.o
a'.o
T IME Ins)
Fig. 10. Comparison of waveform- and jitter-DLs for U.T. (250
Hz). The upper plot shows the gaussion pulse—interval distribution function (standard deviation = 0.77 ms) and the interval
histogram of a stimulus pulse train generated according to this
function. The lower plot shows a trapezoidal waveform (rise
time/period — 0.02, thick line) and a rectangular waveform
(rise time/period — 0.0, thin line) of a 250 Hz stimulus.
•
53
form DL was 0.03; a trapezoid with a rise time of
0.06 ms was discriminated. The jitter DL was 0.27
ms, meaning that a stimulus with 95% of interpulse intervals between 1.47 and 2.53 ms (mean ± 2
S.D.) could be distinguished from an ordinary
pulse train. Thus, waveform differences comprising less than 0.1 ms could be discriminated,
while pulse trains with jitter required about 1 ms
before being discriminated from ordinary pulse
trains. The temporal differences among discriminable waveforms would be even less if we
considered their charge (integrated current) waveforms (see Fig. 3).
It should be noted again that pulse trains with
jitter were heard as rough, whereas the difference
between square and triangular waves was heard
along a sharp-dull dimension. Pulsatile stimuli (of
short duration, constant amplitude and with interpulse intervals greater than the refractory period)
evoke synchronous activity for both regular and
irregular pulse trains in the whole nerve fiber
population whereas analog waveforms produce
temporally more dispersed discharge patterns. Our
results seem to indicate that the CANS interprets
synchronous interneural timing differences between successive stimulus periods differently than
periodically repeated interneural timing differences within the same cycle.
Several of the questions posed in the Introduction have been answered, at least in part:
1. How well can the CANS use purely temporal
information for frequency discrimination? Stochastic pulse trains permitted us to present
stimuli differing only in fine time structure. For
probabilities > 0.5, performance was as good
as for ordinary pulse trains. Period differences
as small as 0.1 ms could be discriminated.
2. Would a stochastic, low-mean-rate pulse code
be useful for electrical stimulation of the
cochlea? Since frequency discrimination was
best for high-probability pulse trains, which
were least like natural auditory-nerve firing patterns, the answer seems to be `no'. Apparently
the CANS does not ordinarily infer frequency
by analyzing single-unit interval histograms, but
by analyzing the aggregate response of an array
of units, at least in the frequency range ( < 1000
Hz) studied. However, under conditions of
pulsatile electrical stimulation, discrimination
of interpulse periods does not depend on interneural timing differences.
3. What are the limits of waveform discrimination
for analog coding? Rise-time differences (square
wave vs. trapezoid) of less than 0.1 ms could be
discriminated; low-pass filtering, even at 2000
Hz, would eliminate these cues.
4. What are the mechanisms for waveform discrimination? If the CANS were able to distinguish waveforms on the basis of the degree
of synchrony in single-unit period histograms,
subjects should perform as well in detection of
jitter in pulse trains as they do in waveform
discrimination. Since jitter detection was an
order of magnitude worse than waveform discrimination, we conclude that the degree of
synchrony in a neural array is the most important cue.
To interpret the results of both the stochastic
pulse and waveform discrimination experiments,
the notion of a neural array is necessary. Even a
single electrode can be considered to control a
multi-channel' neural receiver. This has implications for implant coding strategies: pulse coding in
which only the timing and amplitude of pulses is
varied probably could not exploit this perceptual
ability while analog codes (or pulse codes in which
pulse duration varies) could do so. The choice of a
coding scheme, however, depends on many factors. Our group has previously recognized the fact
that simpler analog codes (e.g., that used by the
Vienna group) may have advantages over pulse
codes in that the information processing required
for a pulse code inevitably results in a loss of
information, and we are very far from knowing the
`optimal' pulse code [3]. However, analog schemes
are themselves sometimes difficult to use in practice, requiring intricate frequency equali7ation adjustments. The choice of a pulse code by our group
was also based on safety considerations.
The experiments reported here make some sense
in terms of what is known about single unit responses to electrical stimulation, but they raise
many new questions which should be answerable
by single-unit neurophysiologic methods. Do individual VIIIth nerve units possess the degree of
threshold variability which would permit them to
signal the difference between a square and triangular wave by the degree of synchrony in their
54
interval histograms? (We doubt it.) What are the
interval histograms of such units to high frequencies (1-4 kHz) like? Are they still unimodal? At
the cochlear nucleus level, are units responding
preferentially to specific pulse intervals to be
found? (We expect they will be found.)
Psychophysical work also needs to be done.
Simple frequency discrimination above 500 Hz for
electrical stimulation is admittedly poor, but the
results reported here give some intriguing hints
that the CANS is capable of some rather fine
temporal discrimination — on the order of less
than 0.1 ms. With regard to consonant discrimination, it would be particularly interesting to know
how well electrically stimulated patients can perform on tasks involving rapid frequency changes.
We wish to emphasize that the results reported
are only a minimum estimate of what at least some
patients can do. There is, of course, tremendous
variability in the abilities of implant patients, depending on (presumably) nerve survival, intellectual ability, etc. Many authors have considered that
the differences among patients would assist in
choosing between single-channel and multi-channel devices.
These differences may also be relevant to a
choice among different coding strategies for particular patients receiving single-channel devices:
a patient with very limited neural survival might
best be served by a very limited pulse code which
minimizes problems of dynamic range, while
another patient with a more generous array of
neurons might profit from an analog coding scheme
permitting him to exploit his `multi-channel' auditory nerve receiver.
Acknowledgements
This work was supported by Swiss National
Research Foundation Grant No. 3.848.0.79. and
by a Teacher-Investigator Development Award
from the National Institute for Neurological and
Communicative Disorders and Stroke (to R.A.D.).
The authors would like to acknowledge the critical
comments of Drs. C. Van den Honert, P. Stypulkowski, I. and E. Hochmair and R. Hartmann.
References
1 Bilger, R.C. (1977): Psychoacoustic evaluation of present
prostheses. Ann. Otol. Rhinol. Laryngol. 86 (Suppl. 38),
92-140.
2 Dillier, N., Spillmann T. and Guentensperger, J. (1983):
Computerized testing of signal encoding strategies with
round window implants. In: Cochlear Prostheses: An International Symposium. Editors: C.W. Parkins and J.W.
Anderson. Ann. N.Y. Acad. Sei. 405, 360-369.
3 Dillier, N. and Spillmann, T. (1984): Signal coding for
single channel stimulation. In: Cochlear Implants, pp.
233-242. Editors: R.A. Schindler and M.M. Merzenich.
Raven Press, New York.
4 Evans, E.F. (1978): Place and time coding of frequency in
the peripheral auditory system: Some physiological pros
and cons. Audiology 17, 369-420.
5 Godfrey, D.A., Kiang, N.Y.-S. and Norris, B.E. (1975):
Single unit activity in the posteroventral cochlear nucleus of
the cat. J. Comp. Neurol. 162, 247-268.
6 Goldberg, J.M. and Brown, P.B. (1969): Responses of
binaural neurons of dog superior olivary complex to dichotic tonal stimuli: some physiological mechanisms of
sound localization. J. Neurophysiol. 32, 613-636.
7 Goldstein, J.L. (1978): Mechanisms of signal analysis and
pattern perception in periodicity pitch. Audiology 17,
421-445.
8 Hartmann, R., Topp, G. and Klinke, R. (1984): Electrical
stimulation of the cat cochlea - discharge pattem of single
auditory fibers. Adv. Audiol. 1, 18-29.
9 Hartmann, R., Topp, G. and Klinke, R. (1984): Discharge
pattems of cat primary auditory fibers with electrical stimulation of the cochlea. Hearing Res. 13, 47-62.
10 Hochmair-Desoyer, IJ., Hochmair, ES., Burian, K. and
Fischer, R.E. (1981): Four years of experience with cochlear
prostheses. Med. Progr. Technol. 8, 107-119.
11 Hochmair, E.S. and Hochmair-Desoyer, I.J. (1983): Percepts elicited by different speech-coding strategies. In:
Cochlear Prostheses: An International Symposium. Editors:
C.W. Parkins and J.W. Anderson. Ann. N.Y. Acad. Sci.
405, 268-279.
12 Hochmair-Desoyer, IJ., Hochmaier, ES., Burian, K. and
Stiglbrunner, H.K. (1983): Percepts from the Vienna
cochlear prosthesis. In: Cochlear Prostheses: An International Symposium. Editors: C.W. Parkins and J.W. Anderson. Ann. N.Y. Acad. Sei. 405, 295-306.
13 Hochmair-Desoyer, IJ., Hochmair, E.S. and Stiglbrunner,
H.K. (1984): Psychoacoustic temporal processing and Speech
understanding in cochlear implant patients. In: Cochlear
Implants, pp. 291-394. Editors: R.A. Schindler and M.M.
Merzenich. Raven Press, New York.
14 Hochmair, E.S. and Hochmair-Desoyer, IJ.: Aspects of
sound signal processing using the Vienna intra- and extracochlear implants. In: Cochlear Implants pp. 101-110.
Raven Press, New York.
15 House, W.F. and Berliner, K.I. (1982): Cochlear implants:
Progress and perspectives. Ann. Otol. 91, Suppl. 91.
16 Johnson, D.H. (1980): The relationship between spike rate
•
55
and synchrony in responses of auditory nerve fibers to
single tones. J. Acoust. Soc. Am. 68, 1115-1122.
17 Kiang, N.Y.-S. and Moxon, E.C. (1972): Physiological considerations in artificial stimulation of the inner ear. Ann.
Otol. RhinoL Laryngol. 81, 714-730.
18 Langner, G. (1981) Neuronal mechanisms for pitch analysis
in the time domain. Exp. Brain Res. 44, 450-454.
19 Levitt, H. (1980): Adaptive testing in audiology. Scand.
Audiol. Suppl. 6, 241-291.
20 Licklider, J.C.R. (1959): Three auditory theories. In: From
Psychology: A Study of a Science, Vol. 1, pp. 41-144.
Editor: S. Koch. McGraw-Hill, New York.
21 Loeb, G.E., White, M.W. and Merzenich, M.M. (1983):
Spatial cross-correlation. A proposed mechanism for acoustic pitch perception. Biol. Cybern. 47, 149-163.
22 Miller, M.I. and Sachs, M.B. (1984): Representation of
voice pitch in discharge patterns of auditory nerve fibers.
Hearing Res. 14, 257-279.
23 Muller, C.G. (1983): Comparison of percepts found with
cochlear implant devices. In: Cochlear Prostheses: An International Symposium. Editors: C.W. Parkins and J.W.
Anderson. Ann. N.Y. Acad. Sei. 405, 412-420.
24 Rose, J.E., Brugge, J.F., Anderson, D.J. and Hind, J.E.
(1967): Phase-locked response to low-frequency tones in
single auditory nerve fibers of the squirrel monkey. J.
Neurophysiol. 40, 769-793.
25 Sachs, M.B., Young, E.D. and Miller, M.I. (1983): Speech
encoding in the auditory nerve: Implications for cochlear
implants. In: Cochlear Prostheses: An International Symposium. Editors: C.W. Parkins and J.W. Anderson. Ann.
N.Y. Acad. Sei. 405, 94-113.
26 Sachs, M.B. (1983): Round table discussion. In: Cochlear
Prostheses: An International Symposium. Editors: C.W.
Parkins and J.W. Anderson. Ann. N.Y. Acad. Sei. 405,
510-511.
27 Simmons, F.B. (1983): Perceptions from modiolar (eighth
nerve) stimulation. In: Cochlear Prostheses: An International Symposium. Editors: C.W. Parkins and J.W. Anderson. Ann. N.Y. Acad. Sei. 405, 259-263.
28 Spillmann, T., Dillier, N. and Guentensperger, J. (1982):
Electrical stimulation of hearing by implanted cochlear
electrodes in humans. Appl. Neurophysiol. 45, 32-37.
29 Spillmann, T. and Dillier, N. (1984): Round window
cochlear implant. In: Cochlear Implants, pp. 157-165. Editors: R.A. Schindler and M.M. Merzenich. Raven Press,
New York.
30 Van den Honert, C. and Stypulkowski, P.H. (1984): Physiological properties of the electrically stimulated auditory
nerve. II. Single fiber recordings. Hearing Res. 14, 225-243.
31 Wever, E.G. (1949): Theory of Hearing. Princeton University Press, Princeton, NJ.