A Corpus-Based Approach to Speech Enhancement From Nonstationary Noise

An iterative longest matching segment approach to speech enhancement with additive noise and channel distortion

“…This work is an extension of our previous work described in Ming et al (2011Ming et al ( , 2013. In Ming 11 et al (2011), we described a corpus-based approach for speech enhancement from additive noise.…”

“…First, we extend the approach to single-channel speech enhancement with both additive 15 noise and channel distortion (i.e., convolutional noise). Second, in Ming et al (2011) we modeled 16 unknown noise using a combination of multicondition model training and missing-data decoding; in 17 this extended research we present an improved method to model noise for speech estimation, which 18 shares some characteristics with the speech separation method described in Ming et al (2013). 19 Finally, we further extend the single-pass estimation algorithm to an iterative estimation algorithm; 20 the new algorithm uses the previous corpus-based noise and channel estimates to update the corpus 21 speech model for improved speech estimates.…”

“…This 13 inequality can be proved conveniently (see Ming et al, 2011Ming et al, , 2013. Based on (7), therefore, we 14 can obtain an estimate of the longest noisy segment Y t:τ from t with stationary noise and with a 15 matched corpus speech segment, through maximizing the posterior probability P (λ S ζ:η |Y t:τ ) with 16 respect to τ and the corpus segment candidate S ζ:η .…”

“…2 shows the same situation for the estimation of the 5 underlying noise frames). We can obtain an estimate of the underlying speech frame at t by using of the matched corpus speech frame for each noisy frame, we can reconstruct the underlying clean 13 speech frame by forming a Wiener filter as in (Ming et al, 2011). However, we found that this 14 may not be the best method for speech recognition, because of the likely mismatch of the enhanced 15 speech features against the training data.…”

“…For speech enhancement, while Wiener filtering based on the matched corpus speech frame, as 23 described in (Ming et al, 2011), can be used to suppress the additive noise, it is not effective for 24 recovering speech from channel distortion. Therefore, in our speech enhancement experiments with 25 the Aurora 4 database, we reconstruct the waveform for each underlying speech frame by using 26 the magnitude spectrum of the matched corpus speech frame.…”