雑音抑圧とSPLICEを組み合わせた雑音環境下音声認識

摘要

近年携帯端末や力ーナビゲーシヨンシステムなどを，音声入力を用いて操作する場面が多くなつている.しかし実環境下での音声入力は様々な雑音が混入してしまい，クリーン環境で構築された音響モデルのみが使われていた場合，認識率の大きな低下を招くことになる.音声の特徴量から雑音の影響を取リ除く手法は特徴量正規化，特徴強調など様々な手法が提案されているが，単一の手法で影響をすべて取リ除くことは難しい.そこで本稿ではこれらの手法を組み合わせてよリ雑音に頑健な音声認識を目指す.SPLICE(Stereo-Piecewise LInear Compensation for Environments)は任意の特徴量を入力として扱える高い汎用性を持っている.そこで既存手法とSPLICEの組み合わせの中で，各手法の特性が有効に働くを組み合わせを検討する.Aurora-2データベースを用いた英語連続数字音声認識のタスクにおいて，AFE (Advanced Front-End), SPLICE, HEQ (Histogram EQualization)を順番に適用した特徴量がどの環境においても優れた性能を示した.%Recently, voices are often used as input modality to mobile terminals or car navigation systems. However, the voices recorded in actual environment frequently contain various kinds of noise and, because of that, the performance of automatic speech recognition systems is largely degraded. To reduce the influence of noise, various methods such as feature normalization or feature enhancement are proposed but a single method cannot reduce all types of the influence of noise. In this paper, we aim at noise robust recognition to combine those methods. SPLICE (Stereo-Piecewise Linear Compensation for Environments) has the characteristics to accept any type of features as its input. So, we investigate a good combination of the existing methods and SPLICE to maximize the performance and the robustness. In English connected digit recognition experiments performed on the Auroral-2 database, the combination of AFE, SPLICE and HEQ in this order showed the highest performance and robustness.