Speaker adaptation algorithms based on piecewise-moving adaptive segment quantization method

Yoshianao Shiraki, Masaaki Honda

Research output: Contribution to journalArticle

Abstract

A speaker adaptation algorithm based on piecewise movement in the speech spectrum space is proposed for speaker-independent, very-low-bit-rate coding. The universal codebook, which is constructed beforehand and is based on the features of the spectral distribution of an unknown speaker, is adapted to the speaker. The adaptation is based on a process which moves the center of gravity of the universal codebook for several clusters. Two algorithms are proposed for the adaptation, simplified method and a minimum distortion method. The minimum distortion method theoretically ensures that quantization distortion does not increase for the speech data for adaptive training. A coding experiment using spectral segments of actual speech showed that the method has a higher adaptation speed compared than the traditional LBG method. It was also shown that the minimum distortion method converges to the same distortion as the LBG method.

Original languageEnglish
Pages (from-to)58-66
Number of pages9
JournalSystems and Computers in Japan
Volume21
Issue number11
Publication statusPublished - 1990
Externally publishedYes

    Fingerprint

ASJC Scopus subject areas

  • Computational Theory and Mathematics
  • Hardware and Architecture
  • Information Systems
  • Theoretical Computer Science

Cite this