Dynamic frequency warping for speaker adaptation in automatic speech recognition

Abstract

Variation in speech characteristics due to speaker differences is a major source of error in the automatic speech recognition of an arbitrary speaker. In the present paper, a nonlinear spectral normalization procedure for speaker adaptation is proposed. This procedure employs the dynamic programming algorithm for warping the log-power spectrum along the frequency axis and does not require prior knowledge about the input speaker. It is shown that the dynamic frequency warping approach compensates successfully for speaker differences in spectral parameters. This approach is applied to the recognition of vowels and the isolated words of an arbitrary speaker. Six different types of dynamic programming algorithms were studied for this purpose over a range of warping-window lengths. It is shown that dynamic frequency warping does not improve the performance of speaker-independent speech recognition systems.