雑音低減や耐雑音音声認識へのアプローチにスペクトル強調と特徴量補正があり，それぞれ固有の利点をもつ．スペクトル強調は，スペクトル包絡と調波構造の両方を考慮した高解像度の音声のモデルを用いることができるため，高い雑音低減性能をもつ．一方，特徴量補正は，低次元の特徴量空間で表されたスペクトル包絡のモデルを用いるため，スペクトルの微細構造の違いに頑健である．これら双方の利点を享受するために，本稿ではまず，スペクトル包絡と調波構造のモデルを用いたスペクトル強調方法を提案し，さらにこれを特徴量補正と統合する．実験の結果，提案法は，音声認識精度とSN比の改善の両方において高い性能を示した．%Spectrum enhancement and feature compensation are typical approaches to noise reduction and noise robust speech recognition, and they have different advantages. The spectrum enhancement approach can potentially achieve a higher noise reduction rate by leveraging a high-resolution speech model that takes account of both a specral envelope and a harmonic structure. On the other hand, the feature compensation approach is superior in terms of robustness against change in spectral fine structures due to the low dimensionality of acoustic features used typically. The goal of this study is to develop a method that integrates these two approaches so that we can reap the benefits of both. Furthermore, in the course of the derivation of the proposed method, also proposed is a novel spectrum enhancement method, which uses models of a spectral envelope and a harmonic structure. Experimental results show that the proposed method improves both a signal to noise ratio and a speech recognition score significantly.
展开▼
