音声認識の数理モデル

概要

論文の詳細を見る
This paper describes mathematical models widely used in speech recognition systems, including linear predictive analysis, cepstral analysis, dynamic time warping (DTW), the stochastic models of sentence speech recognition systems, hidden Markov models (HMMs), and stochastic language models. Stochastic models such as HMMs and stochastic language models are now used in most advanced speech recognition systems, because they are powerful and useful if large speech and language databases are available for training and if the characteristics of those databases are maintained in the utterances to be recognized. However, in many cases, speakers, microphones, transmission systems, and additive noise usually differ between the training samples and the utterances to be recognized. Even if they are the same, manner of speaking changes with the speaker. Language characteristics are also variable. For example, words and their sequences that were not included in the language database often appear in the utterances to be recognized. Therefore, robustness and adaptability are very important characteristics for these stochastic models.
日本応用数理学会の論文
1993-09-16