A Study of Speech Enhancement Approach and Model Adaptation Approach for Speech Recognition

• SHOZAKAI Makoto

  Asahi Chemical Industry Co., Ltd./Nara Institute of Science and Technology

• NAKAMURA Satoshi

  Nara Institute of Science and Technology

• SHIKANO Kiyohiro

  Nara Institute of Science and Technology

Bibliographic Information

Other Title

• 音声認識における音声強調手法及びモデル適応化手法の検討

Abstract

This paper an investigation of speech enhancement and model adaption for speech recognition. As a result of comparison study of Spectral Subtraction, Continuous Spectral Subtraction and Minimum Mean Square Error Estimator, we show CSS performs best as speech enhancement approach for additive noise. Then we compare several variations of Cepstrum Mean Normalization for multiplicative distortion to suggest the best performance may be obtained when two separate mean cepstrum vectors for speech and non-speech are calculated every speaker. Finally exemplified are a couple of recognition tasks for which speech enhancement is superior to model adaptation and vice versa, after describing an approximation method of HMM composition.

Journal

• IEICE technical report. Speech

  IEICE technical report. Speech 96 (92), 53-60, 1996-06-13

  The Institute of Electronics, Information and Communication Engineers

Keywords

• speech recognition
• speech enhancement
• model adaptation
• additive noise
• multiplicative distortion

Details

• CRID: 1570572702386947840
• NII Article ID: 110003296361
• NII Book ID: AN10013221
• Text Lang: ja
• Data Source:
  • CiNii Articles