Separating Speech From Noise Challenge

We have used the data from the PASCAL CHiME challenge with the goal of training a Support Vector Machine (SVM) to estimate a noise mask that labels time-frames/frequency-bins of the audio as 'reliable' or 'unreliable'. This noise mask could be used by another block in the signal processing pipeline to treat the unreliable data as missing and then replace the missing data with an estimate of the clean audio by searching a corpus of clean audio samples for the most probable match using the features of the unreliable data and the surrounding audio. For this project, we have focused on the noise mask estimation using an SVM and not on the data imputation portion of the problem. It has been demonstrated by Kallasjoki et al. [1] that given a good noise mask, it is possible to achieve significant improvements in automated speech recognition accuracy rates by replacing unreliable portions of the audio with estimates of the clean audio. In order to judge the SVM classification accuracy in generating a noise mask, we needed an 'oracle mask' which gave the correct answer for the mask. The oracle mask generates a label of reliable or unreliable for time frame/frequency pairs, using the Mel filterbank energies of the clean signal and the noisy signal and labeling time/frequency pairs as unreliable if the SNR is less than-3 dB. Using the oracle mask, we estimated the best-case performance of noise mask estimation and data imputation by replacing unreliable time/frequency segments (as labeled by the oracle mask) with the known clean speech audio. As expected, this achieves very good recognition rates that approach the accuracy of the using the clean speech audio. As previously discussed, previous work [1] has demonstrated significant performance improvements using an oracle mask along with sparse imputation methods. Estimating the noise mask was a weak point in the paper cited above. Our goal has been to improve on the noise mask generation using machine learning methods. Since the automatic speech recognition system used in the CHiME project (HTK) uses the Mel filterbank energies as features for a hidden Markov model, we started by using these same filterbank energies as features for the support vector machine to estimate a noise mask. We have used the freely available tools LIBLINEAR [2] and LIBSVM [3] for training and prediction rather than writing an SVM from scratch, allowing us to focus on feature selection, kernel selection, …