Research on Frequency Envelope Mapping Based on GMM

Abstract

A mapping function between different speaker frequency envelopes based on GMM is put forward, in which several linear functions are combined by probability, and LSF parameter is adopted to reflect the frequency envelope. The initial mapping parameter are gotten by the relation of average and covariance between source and target GMM. Furthermore, the EM algorithm is adopted to refine the mapping function, by which the estimation of target feature can be gotten from the source feature. All the analysis of histogram, the difference between estimation and the real target feature with time and the DTW distance show that, the proposed mapping function can well characterize the relationship between source and target features.