話者クラス音響モデルを用いた講演音声認識の性能向上(音響モデル,認識,理解,対話,一般)  [in Japanese] Improvement of lecture speech recognition by using speaker-class models  [in Japanese]

本稿では講演音声認識の性能向上を目指し,話者クラス音響モデルの検討を行った.話者クラスモデルの使用法として,1)尤度基準によるモデルの自動選択,2)システム統合,の検討を行った.さらに,この認識結果を利用して教師なし適応の性能向上の検討を行った.以上の評価を日本語話し言葉コーパスを用いて行った.認識実験の結果,ベースラインの単語誤り率19.75%に対し,話者クラスモデルの自動選択で19.11%,システム統合で18.65%を得た.また,一般的なMLLR適応で17.50%,話者クラス音響モデルを利用した適応で17.03%,適応後の話者クラス音響モデルの出力統合により16.79%を得た.以上より,講演音声認識において,提案手法が有効であることが分かった.

This paper describes a new method based on speaker-class (SC) models in order to improve the performance of lecture speech recognition. We investigate two usages of SC models: 1) the automatic selection of SC model by likelihood basis, and 2) the system combination of SC models. Furthermore, unsupervised speaker adaptation is studied by using SC models. The evaluation was conducted on CSJ (Corpus of Spontaneous Japanese). As the results, a word error rate of 19.11% was obtained by using the automatic selection method, and 18.65% was obtained by using the system combination, while 19.75% was obtained in the baseline experiment. In addition, 17.03% was obtained by using the adaptation method based on SC models, and 16.79% was obtained by using the system combination based on adapted SC models, while 17.50% was obtained by using conventional MLLR. The results showed that the proposed methods were effective for lecture speech recognition.