5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Robust Speech Recognition Using Discriminative Stream Weighting and Parameter Interpolation

Stephen M. Chu, Yunxin Zhao

Beckman Institute and Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, USA

This paper presents a method to improve the robustness of speech recognition in noisy conditions. It has been shown that using dynamic features in addition to static features can improve the noise robustness of speech recognizers. In this work we show that in a continuous-density Hidden Markov Model (HMM) based speech recognition system, weighting the contribution of the dynamic features according to SNR levels can further improve the performance, and we propose a two-step scheme to adapt the weights for a given Signal to Noise Ratio (SNR). The first step is to obtain the optimal weights for a set of selected SNR levels by discriminative training. The Generalized Probabilistic Decent (GPD) framework is used in our experiments. The second step is to interpolate the set of SNR-specific weights obtained in step one for a new SNR condition. Experimental results obtained by the proposed technique is encouraging. Evaluation using speaker independent digits with added white Gaussian noise shows significant reduction in error rate at various SNR levels.

Full Paper

Bibliographic reference.  Chu, Stephen M. / Zhao, Yunxin (1998): "Robust speech recognition using discriminative stream weighting and parameter interpolation", In ICSLP-1998, paper 0690.