Speaker adaptation algorithms based on piecewise moving adaptive segment quantization method

Abstract

This paper presents a new speaker adaptation method based on a piece-wise linear mapping of spectral code vector space into thespectral vector space of an unknown speaker. Adaptation is performed by modifying the code vectors to give a better fit to input the spectral vectors, while maintaining the local continuity of distribution of the original code vectors. Two adaptation algorithms are presented. One is a minimum distortion method which provides a monotonic non-increasing distortion for the training data. The other is a simplified method. These algorithms are investigated in terms of distortion performance and adaptation speed. Experimental results show that the proposed method provides much higher adaptation speed than the LBG algorithm and that the distortion performance approaches that of the LBG algorithm as the size of training data increases.

Speaker adaptation method for fenonic markov model‐based speech recognition

“…This method generally requires a large number of training speech data to construct an accurate codebook. An adaptation without a supervisor has been proposed [7] which does not utilize the time-series data such as alignment but does utilize the information concerning the feature parameter distribution. A more accurate recognition is expected along this direction by extending the method to the training with a supervisor.…”