Improved Speaker Markov Modelling for Unsupervised Speaker Normalization

Abstract

We propose new methods of improved speech recognition with speaker-variable information. Hidden Markov Model-based recognizers which are trained by reference speaker(s) (RS) are normalized by our two different approaches to give a better speaker-independent recognition rate. Our normalization methods are based on the same principle of inter-speaker Markov mapping. This mapping gives inter-speaker parameters which are used differently in our two approaches. The first Speaker Markov Model Converter (SMMC) converts new speaker spectral data into label data similar to that of the reference speaker utterance, which is passed directly to the recognizer. In the second Integrated Markov Model (IMM) approach, inter-speaker emission probabilities (ISE) are integrated as weights to the HMM emission probabilities. The recognizer in this case is modified according to interspeaker variable information whereas the normalization is done in context. The inter-speaker mapping in both cases are unsupervised to save new speaker (NS) effort. HMM score thresholding, template matching and DP thresholding techniques are applied to select suitable data for unsupervised mapping of NS and RS data. This mapping is done in parallel to the recognition process. Iterations are performed to improve the unsupervised mapping.