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

TLDR: This paper presents a new approach to speech enhancement from single-channel measurements involving both noise and channel distortion (i.e., convolutional noise), and demonstrates its applications for robust speech recognition and for improving noisy speech quality.

About: This article is published in Computer Speech & Language.The article was published on 2014-11-01 and is currently open access. It has received 8 citations till now. The article focuses on the topics: Speech enhancement & Voice activity detection.

Figures

Figure 1: Illustration of the proposed longest matching segment (LMS) approach. Assume that shown on the top of each section is a noisy signal power spectral density (PSD) sequence for a specific frequency bin k sampled at consecutive discrete frame times t. The bottom of each section shows the combination of some corpus signal PSD segment and stationary noise PSD segment to match a noisy signal PSD segment assuming stationary measurement noise in the segment. The longer the matched segments found, the more specific the matched corpus signal, subject to the nonnegative, constant noise PSD constraint.

Figure 2: A smoothed mean estimate (dashed curve) of the noise power variation (solid curve) by averaging the segment-based stationary noise mean estimates (horizontal straight lines).

Table 6: WER of the baseline recognition system for the enhanced speech, comparing the LMS algorithm with other speech enhancement algorithms, for the Aurora 4 data with noise (test condition B).

Figure 6: Histogram of the length (in number of frames) of the longest matched segments found by the LMS algorithm as a function of the iteration index, for the test speech with combined noise and channel distortion.

Table 1: Summary of the databases used in the experiments.

Table 2: Aurora 4 word error rates (WER) for the unprocessed noisy speech and the reconstructed speech features from the LMS enhancement system, based on a HTK baseline speech recognition system trained using wideband, clean training data, for different test noise conditions with and without channel distortion.

Frequently asked questions

Q1. What are the contributions mentioned in the paper "An iterative longest matching segment approach to speech enhancement with additive noise and channel distortion" ?

6 This paper presents a new approach to speech enhancement from single-channel measurements involving both noise and channel distortion ( i. e., convolutional noise ), and demonstrates its applications for robust speech recognition and for improving noisy speech quality. Second, the authors present an improved method for modeling noise for speech estimation. Third, the authors present an iterative algorithm which updates the noise and channel estimates of the corpus data model. In experiments using speech recognition as a test with the Aurora 4 database, the use of their enhancement approach as a preprocessor for feature extraction significantly improved the performance of a baseline recognition system. 

Q2. How is the likelihood function of Yt: determined?

The likelihood of the match between the two segments Yt:τ and 17 Sζ:η is decided through optimizing the parameters g, qk and hk on the segment level assuming 18 stationary noise and constant channel characteristic in the segment. 

Q3. What is the likelihood function of Yt: associated with S:?

In other words, given a noisy 19 segment Yt:τ , p(Yt:τ |λSζ:η ) indicates the likelihood of the noisy segment with stationary noise and 20 with an accordingly matched corpus segment Sζ:η, subject to a time-invariant channel factor. 

Q4. How do the authors add compensation into the corpus speech model?

By introducing channel and noise 34 compensation into the corpus GMM, the authors therefore introduce the compensation into all the corpus 35utterances built on the GMM used for finding the matched segments. 

Q5. What is the noise model used to calculate the likelihood of the measurement?

This new noise model was used as an alternative to the white noise model - 11 in calculating the likelihood of the measurement in (3) and (5), the noise model of the two which 12 produced a larger likelihood would be used. 

Q6. How long does the LMS approach take to find the longest noisy segment?

the authors propose the longest matching segment (LMS) approach: at each time t, the authors 11 find the longest noisy segment from t that can assume stationary noise and has an accordingly 12 matched corpus speech segment, subject to a constant channel factor. 

Q7. What is the difference between the smoothed channel and noise estimate?

The smoothed channel and noise 27 estimates can be used to modify the wideband, clean corpus speech model to reduce the mismatch 28 against the noisy measurement, or used to reduce the level of distortion in the noisy measurement, 29 thereby reducing the error in segment matching. 

Q8. What is the effect of the speech 27 gain resolution on the search?

This is subject to the constraint that the power of the model 26 of speech plus noise should not exceed the power of the noisy measurement; the use of the speech 27 gain resolution to quantize the noise gain range for the search reduces the amount of computation 28 for (3) and (5). 

Q9. How many utterances are in the second part of the test data?

Like the first part of test data, the second 5 part of test data includes six test sets with both noise and channel distortion, plus one test set 6 without noise and with channel distortion only; each test set contains 330 utterances.