Experiments in audio source separation with one sensor for robust speech recognition

This paper focuses on the problem of noise compensation in speech signals for robust speech recognition. We investigate on a novel paradigm based on source separation techniques to remove music from speech, a common situation in broadcast news transcription tasks. The two methods proposed, namely adaptive Wiener filtering and adaptive shrinkage, rely on the use of a dictionary of spectral shapes to deal with the non-stationarity of the signals. Unlike most classical noise suppression methods, we assume a prior knowledge of the sources that are mixed. The proposed algorithms are compared to simple standard approaches on the source separation task and assessed in terms of average distortion. Their effect on the entire transcription system is eventually compared in terms of word error rate. Results indicate that source separation techniques show some effectiveness for robust transcription at signal/noise ratio lower than 15 dB. We also observe that the improvement of the word error rate is correlated to the spectral distortion rather than to specific source separation performance measure such as the signal to interference ratio.