Combining Neural Networks and Hidden Markov Models for Speech Recognition

摘要

Automatic Speech Recognition (ASR) is a challenging classification task over sequences of acoustic features, extracted from the vocal signal, according to some linguistic model. Hidden Markov models (HMM) are the most common and successful tool for ASR, allowing for high recognition performance in a variety of difficult tasks (speaker independent, large vocabulary, continuous speech). In spite of that, standard HMMs suffer from relevant limitations. Artificial neural networks (ANN) have been proposed as an alternative, nonparametric paradigm for ASR since the late Eighties. ANNs were successfully applied in reduced size tasks (e.g., phoneme or isolated word recognition) but they did not succeed in solving the general ASR problem, due to their lack of capability to model long-term dependencies. Combining ANN and HMM within a unifying framework is a suitable approach to ASR, which takes benefit from both techniques and overcomes the corresponding limitations. This paper reviews the main topics concerning hybrid HMM/ANN systems for ASR, summarizing some major architectures of this kind.