Speech Enhancement for ASR by Hans Hwang 8/23/2000 Reference 1. Alan V. Oppenheim,etc., ” Multi-Channel Signal Separation by Decorrelation ”,IEEE Trans. on ASSP, , Yunxin Zhao,etc., ” Adaptive Co-channel Speech Separation and Recognition ”,IEEE Trans. On SAP, , Ing Yang Soon,etc., ” Noisy Speech Enhancement Using Discrete Cosine Transform ”,Speech communication, ,1998
Outline Signal Separation by S-ADF/LMS Speech Enhancement by DCT Residual Signal Reduction Experimental Results
Speech Signal Separation Introduction: -To Recover the desired signal and identify the unknown system from the observation signal -Speech signal recovered from SSS will increase SNR and improve the speech recognition accuracy -Specifically consider the two-channel case
SSS cont ’ d Two-channel model description A and B are cross-coupling effect between channels and we ignore the transfer function of each channel. x i (t) is source signal and y i (t) is acquired signal
SSS (cont ’ d) Source separation system (separate source signals out from acquired signals) and called decoupling filters and modeled as FIR filter
SSS by ADF Calculate the FIR coeff. by adaptive decorre- lation filter(ADF) proposed by A. V. Oppenheim in The objective is to design decoupling filter s.t., the estimated signals are uncorrelated. -The decoupling filtering coeff. ’ s are estimated iteratively based on the previous estimated filter coeff. ’ s and current observations
SSS by ADF (cont ’ d) The closed form of decoupling filters where
SSS by ADF (cont ’ d) Choice of adaptation gain -As time goes to infinite the adaptation gain goes to zero for the system stable consideration. -Optimal choice adaptation gain for the system stability and convergence. -
SSS by ADF (cont ’ d) The experiment of :
Source Signal Detection(SSD) Introduction -If one of the two is inactive then the estimated signals will be poor by ADF and cause the recog- nition errors. -So the ASR and ADF are performed within active region of each target signal.
SSD (cont ’ d)
SSD by coherence function If then
SSD (cont ’ d) - decision variable -Decision Rule:
SSD (cont ’ d) -Implementation using DFT and Result
SSD (cont ’ d)
Improved Filter Estimation Widrow ’ s LMS algorithm proposed in If we don ’ t know A or B in observation(i.e., one of the source signals is inactive) then the estimation of filters will cause much errors compared to the actual filters. -If we know source signal 2 is inactive(using SSD) then we only estimate filter B and remain filter A unchanged.
Improved Filter Estimation LMS algorithm and result
Experimental Results -Evaluate in terms of WRA and SIR
Experimental Result (cont ’ d) *Use 717 TIMIT sentences to train 62 phone units. Front-end feature is PLP and its dynamic. Grammar perplexity is 105. After acoustic normalization
Speech Enhancement using Discrete Cosine Transform Motivation -DCT provides significantly higher compaction as compared to the DFT
SE Using DCT (cont ’ d) -DCT provides higher spectral resolution than DFT -DCT is real transform so it has only binary phases. Its phase won ’ t be changed unless added noise is strong.
Estimating signal by MMSE Intorduction -y(t)=x(t)+n(t) and Y(k)=X(k)+N(k) Assume DCT coeff. ’ s are statistically independent and estimated signal is less diffenent from the original signal. -, by Bayes ’ rule and signal model
MMSE (cont ’ d) Estimating signal source by Decision Directed Estimation(DDE) ( proposed by Ephraim & Malah in ‘ 84 ) = 0.98 in computer simulation
Reduction of Residual Signal Introduction -If the source signal more likely exists then the estimated is more reliable. -two states of inputs H 0 :speech absent H 1 :speech present : modified filter output
Reduction of Residual Signal - where
Experimental Results Measure in Segmental SNR *EMFDETFDETF White noise added Fan noise added
Experimental Results