Auditory Scene Analysis and Automatic Speech Recognition in Adverse Conditions Phil Green Speech and Hearing Research Group, Department of Computer Science, University of Sheffield With thanks to Martin Cooke, Guy Brown, Jon Barker.. HCSNet December 2005 Overview • Visual and Auditory Scene Analysis • ‘Glimpsing’ in Speech Perception • Missing Data ASR • Finding the glimpses • Current Sheffield Work • Dealing with Reverberation • Identifying Musical Instruments • Multisource Decoding • Speech Separation Challenge HCSNet December 2005 Visual Scenes and Auditory Scenes • Objects are opaque • Each spatial pixel images a single object • Object recognition has to cope with occlusion HCSNet December 2005 • Sound is additive • Each time/frequency pixel receives contributions from many sound sources • Sound source recognition apparently requires ‘Glimpsing’ in auditory scenes: the dominance effect (Cooke) Although audio signals add ‘additively’, the occlusion metaphor is a good approximation due to loglike compression in the auditory system HCSNet December 2005 Consequently, most regions in a mixture are dominated by one or other source, leaving very few ambiguous regions, even for a pair of speech signals mixed at 0 dB. Can listeners handle glimpses? HCSNet December 2005 The robustness problem in Automatic Speech Recognition • Current ASR devices cannot Clean speech tolerate additive noise, particularly if it’s unpredictable • Listener’s noise-tolerance is 1 or 2 orders of magnitude better in equivalent conditions (Lippmann 97) +noise • Can glimpsing be used as the basis for robust ASR? Requirements: • Adapt statistical ASR to Missing data incomplete data case Mask (oracle) • Identify the glimpses HCSNet December 2005 Classification with Missing Data A common problem: visual occlusion, sensor failure, transmission losses.. Need to evaluate the likelihood that observation vector x was generated by class C , f(x|C) Assume x has been partitioned into reliable and unreliable parts, (xr,xu) Two approaches: Imputation: estimate xu , then proceed as normal Marginalisation: integrate over possible range of xu Marginalisation is preferable if there is no need to reconstruct x HCSNet December 2005 The Missing Data Likelihood Computation In ASR by Continuous Density HMMS, • State distributions are Gaussian Mixtures with diagonal covariance • The marginal is just the reduced dimensionality distribution • The integral can be approximated by ERFS • This is computed independently for each mixture in the state distribution HCSNet December 2005 Cooke et al 2001 Counter-evidence from bounds reliable unreliable Mean spectrum for class C frequency Observed spectrum x Class C matches the reliable evidence well but there is insufficient energy in the unreliable components HCSNet December 2005 Finding the glimpses Auditory scene analysis identifies spectral regions dominated by a single source • Harmonicity • Common amplitude modulation • Sound source location Local SNR estimates can be used to compensate for predictable noise sources. HCSNet December 2005 Cooke 91 Harmonicity Masks • Only meaningful in voiced segments • Can be combined with SNR masks HCSNet December 2005 Aurora Results (Sept 2001) Barker et al 2001 Average gain over clean baseline under all conditions: 65% HCSNet December 2005 Missing data masks from spatial location Sue Harding, Guy Brown • Cues for spatial location are used to separate a target source from masking sources • Interaural Time Difference from corss-correlation between left and right binaural signals • Interaural Level Difference from ratio of energy in left and right ears • Soft masks • Task: • Target source: male speaker straight ahead • One or two masking sources (also male speakers) at other positions • Added reverberation HCSNet December 2005 60 50 40 30 20 10 Localisation mask, ILD/ITD Localisation mask, ILD only Frequency channel Localisation mask, ITD only Frequency channel 60 50 40 30 20 10 40 60 80 100 120 Time (frames) Oracle ITD only, ILD only, combined ITD and ILD. 20 40 60 80 60 50 40 30 20 10 100 120 20 Time (frames) 40 60 80 100 120 Time (frames) 100 90 80 % Accuracy 20 % Accuracy Frequency channel Missing data masks from spatial location (2) 70 60 50 Best performance is with combined ITD and ILD: HCSNet December 2005 40 30 5 7.5 10 15 20 30 Azimuth of masker (degrees) 40 • MD for reverberant conditions (1) Palomäki, Brown and Barker have applied MD to the problem of room reverberation: • Use spectral normalization to deal with distortion caused by early reflections; • Treat late reverberation as additive noise, and apply standard MD techniques. • Select features which are uncontaminated by reverberation and contain strong speech energy. Approach based on modulation filtering: • Each rate map channel passed through modulation filter • Identify periods with enough energy in the filtered output • Use these to define mask on HCSNet December 2005 original rate map MD for reverberant conditions (2) HCSNet December 2005 80 60 40 HMM-MLP Baseline 20 MD A priori Mask MD Reverb Mask 0 C le 0. an 7s 2. 35 0. m 7s 3. 05 m 1. 2s 1. 2m 2s 3. 05 m 1. 5s 6. 1. 1m 5s 18 .3 m K. J. Palomäki, G. J. Brown and J. Barker (2004) Speech Communication 43 (1-2), pp. 123-142 Recognition accuracy (%) • Recognition of connected digits (Aurora 2) • Reverberated using recorded room impulse responses • Performance comparable with Brian Kingsbury’s hybrid HMM-MLP recognizer 100 T60/source-receiver distance MD for music analysis (1) • Eggink and Brown have used MD techniques to identify concurrent musical instrument sounds • Part of a system for transcribing chamber music • Identify the F0 of the target note, and only keep its harmonics in the MD mask • Uses a GMM classifier for each instrument, trained on isolated tones and short phrases • Tested on tones, phrases and commercial CD HCSNet December 2005 MD for music analysis (2) J. Eggink and G. J. Brown (2003) Proc. ICASSP, Hong Kong, IV, pp. 553-556 J. Eggink and G. J. Brown (2004) Proc. ICASSP, Montreal, V, pp. 217-220 Clarinet Fundamental Frequency (Hz) • Example: duet for flute and clarinet • All instrument tones correctly identified in this example Flute 700 600 500 400 300 200 100 0 20 40 60 80 100 Time (frames) HCSNet December 2005 120 Multisource Decoding Use primitive ASA and local SNR to identify time-frequency regions (fragments) dominated by a single source… i.e. possible segregations S … but NOT to decide what the best segregation is Instead, jointly optimise over the word sequence W and S Decoding algorithm finds best subset of fragments to match speech source Based on missing data techniques – regions hypothesised as nonspeech are missing Barker, Cooke & Ellis 2003 HCSNet December 2005 Multisource decoding algorithm Work forward in time, maintaining a set of alternative decodings – Viterbi searches based on a choice of speech fragments. When new fragment arrives, split decodings - speech or non-speech? When fragment ends, merge decoders which differ in its interpretation. HCSNet December 2005 Multisource Decoding on Aurora HCSNet December 2005 Multisource decoding with a competing speaker Andre Coy and Jon Barker • Utterances of male and female speakers mixed at 0 db • Voiced regions: Soft Harmonicity masks from autocorrelation peaks • Voiceless regions: fragments from ‘image processing’ • Gender-dependent HMMs. • Separate decoding for male & female • 73.7% accuracy on a connected digit task Informing Multisource Decoding – Work in progress Ning Ma, Andre Coy, Phil Green • HMM Duration constraints • Links between fragments – pitch continuity • ‘Speechiness’ HCSNet December 2005 Speech separation challenge Organisers: Martin Cooke (University of Sheffield, UK) , TeWon Lee (UCSD, USA) • see http://www.dcs.shef.ac.uk/~martin • Global comparison of techniques for separating and recognising speech • Special session of Interspeech 2006 in Pittsburgh (USA) from 17-21 September, 2006. • Task- recognise speech from a target talker in the presence of either stationary noise or other speech. • Training and test data supplied. • One signal per mixture (i.e. the task is "single microphone"). • Speech material- simple sentences from the ‘Grid Task’, e.g. “place white at L 3 now" HCSNet December 2005