Blind Separation of Multi-speaker Speech

• TAKEDA Kazuya

  Information Electronics, Graduate School of Engineering, Nagoya University

• TANIGUCHI Tokohiko

  Information Electronics, Graduate School of Engineering, Nagoya University

• KAJITA Shouji

  Information Electronics, Graduate School of Engineering, Nagoya University

• ITAKURA Fumitada

  Information Electronics, Graduate School of Engineering, Nagoya University

Bibliographic Information

Other Title

• Blind Separation 法による複数話者音声の分離

Abstract

A blind separation method, proposed by A. J. Bell et al. , is evaluated by applying to multi-channel recordings of multi-speaker speech. Simulation experiments clarify that 1)the performance under noisy conditions degrades when the SNR of original signal falls under 10dB, and 2)time difference at the mixing process is critical to the performance. Real situation experiments, where two microphones are used for recording two mixed signals of two speech signals, are performed under three different acoustic environments, i. e. anechoic, soundproof and a laboratory room conditions. The resultant performance in SDR (signal to deviation ratio) of, 10, 7 and 3dB, respectively, show that the most serious deviation from the theoritical assumption of the method is existance of rebervaration causing a time difference at mixing process.

Journal

• Technical report of IEICE. EA

  Technical report of IEICE. EA 97 (76), 31-37, 1997-05-30

  The Institute of Electronics, Information and Communication Engineers

Keywords

• blind separation
• mutual information
• reverbaration

Details

• CRID: 1572824502268829824
• NII Article ID: 110003284008
• NII Book ID: AN10164817
• Text Lang: ja
• Data Source:
  • CiNii Articles