Training algorithm of the codebooks

Takehiro Moriya, Satoshi Miki, Kazunori Mano, Jotarou Ikedo

Research output: Contribution to journalArticlepeer-review

Abstract

An iterative training method is proposed for the design of random codebooks and channel code mapping tables for a CELP coder. The method is based on a structured excitation source, including a conjugate structure, pitch period repetition, and integrated mapping to channel code. All training procedures use a closed-loop distortion measure, which minimize perceptually weighted distortion between the input and the synthesized speech. The training improves SNR by around 1 dB and improves subjective quality. The channel code mapping and channel-matched distortion measure increase robustness against channel errors.

Original languageEnglish
Pages (from-to)397-404
Number of pages8
JournalNTT R and D
Volume43
Issue number4
Publication statusPublished - 1994
Externally publishedYes

ASJC Scopus subject areas

  • Electrical and Electronic Engineering

Cite this