拡散音場理論に基づく残響環境下音声認識 [in Japanese] Speech Recognition under Reverberant Environments Based on the Diffuse Sound Field Theory [in Japanese]
Search this Article
Author(s)
Abstract
残響環境下での音声認識性能向上のための残響除去法を提案する.本方式はスペクトルサブトラクション(SS)法により残響を除去する.その際のSS法の引き去り係数は,拡散音場理論に基づき残響時間から算出する.残響時間を発話から推定することで,事前知識なくさまざまな残響環境においてロバストに残響除去を行うことができる.JEIDA-JCSD (B-set)の音声とIPSJ SIG-SLP残響下音声認識評価環境CENSREC-4を用いた音声認識実験により提案法の有効性を示す.
This paper addresses a dereverberation method for improving speech recognition capability under reverberant environments. This method dereverberates speeches using the Spectral Subtraction (SS) method. According to the diffuse sound theory, subtraction coefficients of the SS method are calculated from a reverberation time. The proposed method can dereverberate speeches in various reverberant environments robustly without prior knowledge because it estimates reverberation times from speeches. Recognition experiments with JEIDA-JCSD (B-set) and CENSREC-4 show that the proposed method improves recognition rate in reverberant environments.
Journal
-
- IEICE technical report
-
IEICE technical report 110(56), 19-24, 2010-05-19
The Institute of Electronics, Information and Communication Engineers