Phoneme Recognition based on AF-HMMs with Optimal State Configuration

摘要

認識と合成の双方に共通の調音運動HMM を利用する，ワンモデル音声認識・合成方式の音声認識モジュールに対して性能を改善する研究を行っている。本文では，調音特徴（AF）を用いたHMM の構成方法に焦点をあて，音素認識正解率と精度を改善するための評価実験を行った結果を報告する。種々のHMM 構成を対象に，JANAS コーパスによる評価を通して，AF に基づくアプローチは，特徴抽出器に音素文脈情報を埋め込むことで，monophone model や少ない混合数においても，高い性能を達成できることを明らかにする。%Speeh recognition based on one-model of articulatory movement HMMs that are commonly applied to both speech recognition (SR) and speech synthesis (SS) is described. In a SR module, speaker-invariant HMMs are applied to recognize articulatory feature (AF) sequence. This paper focuses on our approaches in designing an optimal state configuration for accurate phoneme recognizer based on articulatory movement. We consider several strategies to improve Articulatory Feature (AF) based phoneme recognition and compare its performance. In the experiments on Japanese Newspaper Article Sentences (JNAS) utterances, the approach of separating short vowel and long vowel in 5-states HMM triphone models provides higher accuracy compared to other approaches.