Audio source separation with one sensor for robust speech recognition

Laurent Benaroya (1) , Frédéric Bimbot (1) , Guillaume Gravier (1) , Rémi Gribonval (1)

1 METISS - Speech and sound data modeling and processing

In this paper, we address the problem of noise compensation in speech signals for robust speech recognition. Several classical denoising methods in the field of speech and signal processing are compared on speech corrupted by music, which correspond to a frequent situation in broadcast news transcription tasks. We also present two new source separation techniques, namely adaptive Wiener filtering and adaptive shrinkage. These techniques rely on the use of a dictionary of spectral shapes to deal with the non stationarity of the signals. The algorithms are first compared 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 show that the proposed adaptive Wiener filter approach yields a significant improvement of the transcription accuracy at signal/noise ratios greater than 15 dB.

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Laurent Benaroya, Frédéric Bimbot, Guillaume Gravier, Rémi Gribonval. Audio source separation with one sensor for robust speech recognition. ISCA Tutorial and Research Workshop on Non-Linear Speech Processing (NOLISP), IRISA, May 2003, Le Croisic, France. ⟨inria-00576210⟩