Using Source Models in Speech Separation

Transcription

1 Using Source Models in Speech Separation Dan Ellis Laboratory for Recognition and Organization of Speech and Audio Dept. Electrical Eng., Columbia Univ., NY USA 1. Mixtures, Separation, and Models. Monaural Speech Separation 3. Binaural Speech Separation. Conclusions Source Models in Separation - Dan Ellis /17

2 LabROSA Overview Information Extraction Music Machine Learning Recognition Separation Retrieval Speech Environment Signal Processing Source Models in Separation - Dan Ellis /17

3 1. Mixtures, Separation, and Models Sounds rarely occur in isolation.. so analyzing mixtures is a problem.. for humans and machines chan freq / khz mr :57:3 chan 1 chan 3 freq / khz freq / khz level / db mr-11-1-ce+173.wav chan 9 freq / khz time / sec Source Models in Separation - Dan Ellis /17

4 Mixture Organization Scenarios Interactive voice systems human-level understanding is expected Speech prostheses crowds: #1 complaint of hearing aid users Archive analysis identifying and isolating sound events freq / khz 8 6 dpwe :15: level / db time / sec Unmixing/remixing/enhancement Source Models in Separation - Dan Ellis /17

5 Separation vs. Inference Ideal separation is rarely possible many situations where overlaps cannot be removed Overlaps Ambiguity scene analysis = find most reasonable explanation Ambiguity can be expressed probabilistically i.e. posteriors of sources {Si} given observations X: P({S i } X) P(X {S i }) P({S i }) combination physics source models search over {S i }?? Better source models better inference.. learn from examples? Source Models in Separation - Dan Ellis /17

6 Approaches to Separation ICA Multi-channel Fixed filtering Perfect separation maybe! CASA Single-channel Time-var. filter Approximate separation Model-based Any domain Param. search Synthetic output target x interference n mix proc + - n^ x^ mix x+n stft proc mask istft x^ mix x+n param. fit params source models synthesis x^ or combinations... Source Models in Separation - Dan Ellis /17

7 EM for Model-based Separation Expectation-Maximization algorithm for solving partially-unknown problems (only local optimality guaranteed) EM for model-based separation E-step: find distribution of unknowns p(u) given current model parameters Θ and observations x M-step: optimize Θ to maximize fit to x given current p(u) E-step M-step u is... GMM mixture assignment... T-F cell dominance... current phone of voice i... Source Models in Separation - Dan Ellis /17

8 What is a Source Model? Source Model describes signal behavior encapsulates constraints on form of signal (any such constraint can be seen as a model...) A model has parameters model + parameters instance What is not a source model? detail not provided in instance Excitation source g[n] e.g. using phase from original mixture constraints on interaction between sources e.g. independence, clustering attributes n Resonance filter H(e j! )! n Source Models in Separation - Dan Ellis /17

9 . Monaural Speech Separation Cooke & Lee s Speech Separation Challenge short, grammatically-constrained utterances: <command:><color:><preposition:><letter:5><number:1><adverb:> e.g. "bin white by R 8 again" task: report letter + number for white special session at Interspeech 6 Separation or Description? mix separation identify target energy t-f masking + resynthesis ASR speech models words vs. mix identify speech models find best words model words source knowledge Source Models in Separation - Dan Ellis /17

10 Codebook Models Given models for sources, find best (most likely) states for spectra: p(x i 1,i ) = N (x;c i1 + c i,σ) combination model {i 1 (t),i (t)} = argmax i1,i p(x(t) i 1,i ) can include sequential constraints... different domains for combining c and defining E.g. stationary noise: freq / mel bin 8 6 Original speech In speech-shaped noise (mel magsnr =.1 db) VQ inferred states (mel magsnr = 3.6 db) Roweis 1, 3 Kristjannson, 6 inference of source state 1 time / s 1 1 Source Models in Separation - Dan Ellis /17

11 Speech Recognition Models Decode with Factorial HMM i.e. two state sequences, one model for each voice exploit sequence constraints, speaker differences? IBM superhuman Iroquois system Kristjansson, Hershey et al. 6 model model 1 observations / time fewer errors than people for same speaker, level exploit grammar constraints - higher-level dynamics wer / % 1 5 Same Gender Speakers 6 db 3 db db - 3 db - 6 db - 9 db SDL Recognizer No dynamics Acoustic dyn. Grammar dyn. Human Source Models in Separation - Dan Ellis /17

12 Frequency (khz) Speaker-Adapted (SA) Models Factorial HMM needs distinct speakers Mixture: t3_swila_m18_sbar9n Adaptation iteration 1 Adaptation iteration 3 Adaptation iteration 5 SD model separation Time (sec)!!!!!!!!!! acc % (%% '% %- use eigenvoice speaker space iterate estimating voice & separating speech performs midway between speaker-independent (SI) and speaker-dependent (SD) 89::-6,7",1 Source Models in Separation - Dan Ellis /17 Weiss & Ellis 7 ;1*3/, )8 ) )< #*=,/97, SI SA SD -

13 3. Binaural Speech Separation or 3 sources in reverberation assume just ears Tasks: identify positions of sources (and number?) recover source signals Source Models in Separation - Dan Ellis /17

14 Spatial Estimation in Reverb Mandel & Ellis 7 Model interaural spectrum of each source as stationary level and time differences: converge via EM to a(), τ for each source mask is Pr(X(t,ω) dominated by source i) Left Left Right Right ILD ILD IPD IPD Mask Mask Mask Mask dB 8.77dB Source Models in Separation - Dan Ellis /17

15 Spatial Estimation Results Modeling uncertainty improves results tradeoff between constraints & noisiness.5 db. db 8.77 db 1.35 db 9.1 db -.7 db Source Models in Separation - Dan Ellis /17

16 Combining Spatial + Speech Model Interaural parameters give ILDi(ω), ITDi, Pr(X(t, ω) = Si(t, ω)) Speech source model can give Pr(S i (t, ω) is speech signal) Can combine into one big EM framework... E-step M-step u is: Pr(cell from source i) phoneme sequence Θ is: interaural params speaker params Source Models in Separation - Dan Ellis /17

17 Summary & Conclusions Inferring model parameters is very general.. and very difficult, in general Speech models can separate single channels.. better match to individual better results Spatial cues can separate binaural signals.. but account for uncertainty from e.g. reverb EM-type approach can integrate them both Source Models in Separation - Dan Ellis /17