長さの異なるセグメントを組み合わせた音声認識のための特徴ベクトルの構成法

抄録

In this paper, we propose the dual-width windowed segment (DWWS) which consists of a short cepstrum segment and a long Δ cepstrum segment, as a more effective segmental input vector than the conventional one. First, using a criterion for cluster analysis, we evaluate various compositions of feature vectors based on their ability of phoneme separation to show the effectiveness of DWWS. Then we carry out discrete HMM speech recognition experiments to verify the result of evaluation. As a result, it is shown that the DWWS brings on high recognition performance when a categories-dependent codebook is used for vector quantization.