Speech recognition in adverse environments: Improvements to IMELDA.

Abstract

This thesis deals with speech recognition in adverse environments. The primary problem is the mismatch between training and test conditions. Cases of mismatch include the recording channel, acoustic noise and the speaker. Noise shaping and subband filtering are two noise suppression techniques that work by utilizing properties of speech. Dynamic speech analysis and some feature extraction methods are inherently robust to the influence of noise. Linear discriminant analysis (LDA) can be used to combine disparate sets of speech parameters and obtain the optimal set of features. IMELDA (45) is one such method. In this thesis, we analyse the effectiveness of IMELDA under various training and test scenarios. Theoretical results are first derived and substantiated by simulations. It will be shown that LDA provides a form of noise shaping and the root-deconvolution technique is inappropriate for IMELDA. A new algorithm for predicting recognition performance is proposed and verified. Optimal cross-condition recognition is obtained by utilizing samples of noisy test speech in the within-class covariance, in the so-called QNT IMELDA transform. In the event that the noise is stationary, and can be modelled, we derive an equivalent transform by artificially modifying quiet speech samples. This suffices for the simplest instances. For the extreme helicopter case, we show the best approach to be a combination of band-pass filtering and dynamic analysis of the Mel-scale subbands. Unknown channel noise and additive noise are reduced through the respective subband processing algorithms. Finally, practical issues of applying LDA and integrating subband filtering in a speech recognition system are addressed.