Presentation 2014-06-20
Accurate Recognition of Overlapped Speech : High Speed Speech Separation by Spectral Subtraction and Acoustic Model Training using Separated Speeches
Yuto DEKIURA, Tetsuya MATSUMOTO, Yoshinori TAKEUCHI, Hiroaki KUDO, Noboru ONISHI, Norihide KITAOKA, Kazuya TAKEDA,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) The purpose of this study is to recognize overlapped speech more accurately. In order to achieve this, it is necessary to separate overlapped speech. However, recognition accuracy is insufficient due to the distortion and residual noise caused by the processing of separation. Also, the conventional methods based on Independent Component Analysis, take heavy computational cost. In this study, we propose to train acoustic models using distorted speech and to separate overlapped speech by spectral subtraction of estimated speech uttered by the other speakers. In a recognition task of simulated overlapped speech, the proposed method improved recognition accuracy by approximately 6% and reduced 4/5 of the computational cost compared with a conventional method. Also, training acoustic models using separated speech improved recognition accuracy by approximately 20%.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Overlapped Speech Recognition / Speech separation / Speech spectrum estimation / Spectral Subtraction
Paper # SP2014-56,WIT2014-11
Date of Issue

Conference Information
Committee WIT
Conference Date 2014/6/12(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Well-being Information Technology(WIT)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Accurate Recognition of Overlapped Speech : High Speed Speech Separation by Spectral Subtraction and Acoustic Model Training using Separated Speeches
Sub Title (in English)
Keyword(1) Overlapped Speech Recognition
Keyword(2) Speech separation
Keyword(3) Speech spectrum estimation
Keyword(4) Spectral Subtraction
1st Author's Name Yuto DEKIURA
1st Author's Affiliation Nagoya University()
2nd Author's Name Tetsuya MATSUMOTO
2nd Author's Affiliation Nagoya University
3rd Author's Name Yoshinori TAKEUCHI
3rd Author's Affiliation Department of Computer Science, School of Informatics, Daido University
4th Author's Name Hiroaki KUDO
4th Author's Affiliation Nagoya University
5th Author's Name Noboru ONISHI
5th Author's Affiliation Nagoya University
6th Author's Name Norihide KITAOKA
6th Author's Affiliation Nagoya University
7th Author's Name Kazuya TAKEDA
7th Author's Affiliation Nagoya University
Date 2014-06-20
Paper # SP2014-56,WIT2014-11
Volume (vol) vol.114
Number (no) 92
Page pp.pp.-
#Pages 6
Date of Issue