5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Noise Model Selection For Robust Speech Recognition

Laura Docio-Fernandez, Carmen Garcia-Mateo

University of Vigo, Spain

This paper addresses the problem of mismatch between training and testing conditions in a HMM-based speech recognizer. Parallel Model Combination (PMC) has demonstrated to be an efficient technique for reducing the effects of additive noise. In order to apply this technique, a noise HMM must be trained at the recognition phase. Approaches that estimate the noise model based on the Expectation-Maximization (EM) or Baum-Welch algorithms are widely used. In these methods the recorded environmental noise data are used, and their major drawback is that they need a long sequence of noise data to estimate properly the model parameters. In some real life applications the amount of noise data can be too small, so from a practical point of view, the needed amount of noise is a critical parameter which should be as short as possible. We propose a novel method for obtaining a more reliable noise model than training it from scratch by using a short noise sequence.

Full Paper

Bibliographic reference.  Docio-Fernandez, Laura / Garcia-Mateo, Carmen (1998): "Noise model selection for robust speech recognition", In ICSLP-1998, paper 0579.