EUROSPEECH 2003 - INTERSPEECH 2003
8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003

        

High-Likelihood Model based on Reliability Statistics for Robust Combination of Features: Application to Noisy Speech Recognition

Peter Jancovic, Munevver Kokuer, Fionn Murtagh

Queen's University Belfast, U.K.

This paper introduces a novel statistical approach for combination of multiple features, assuming no knowledge about the identity of the noisy features. In a given set of features, some of the features may be dominated by noise. The proposed model deals with the uncertainty about the noisy features by deriving the joint probability of a subset of features with highest probabilities. The core of the model lies in the determination the number of features to be included in the feature-subset - this is estimated based on calculating the reliability of each feature, which is defined as its normalized probability, and evaluating the joint maximal reliability. For the evaluation, we used the TIDIGITS database for connected digit recognition. The utterances were corrupted by various types of additive noise, which resulted the number and identity of the noisy features varied over time (or changed suddenly). The experimental results show that the high-likelihood model achieves recognition performance similar to the one obtained with a full a-priori knowledge about the identity of the noisy features.

Full Paper

Bibliographic reference.  Jancovic, Peter / Kokuer, Munevver / Murtagh, Fionn (2003): "High-likelihood model based on reliability statistics for robust combination of features: application to noisy speech recognition", In EUROSPEECH-2003, 2161-2164.