Hybrid Hidden Markov Model and Artificial Neural Network for Automatic Speech Recognition

Automatic speech recognition (ASR) is an important topic to be performed by a computer system. This paper presents the use of a hybrid hidden Markov model (HMM) and artificial neural networks (ANNs) for automatic speech recognition. The proposed hybrid system for ASR is to take advantage from the properties of both HMM and ANN, improving flexibility and recognition performance. The hybrid ANN/HMM assumes that the output of an ANN is sent to the HMM for ASR. The architecture relies on a probabilistic interpretation of the ANN outputs. Each output unit of the ANN is trained to perform a non-parametric estimate of the posterior probability of a continuous density HMM state given the acoustic observations. After a brief review of HMM and ANN, the paper reports the theoretical aspects and the performance of the proposed hybrid model. Experimental results are listed to demonstrate the potential of this hybrid model.

I. Introduction

Automatic speech recognition (ASR) has been the most investigated topic in speech processing since early 1960s [1]–[3]. In the 1960s, many researchers believed that the incoming new computer technologies would have made ASR rather an easy task. Unfortunately, ASR emerged to be a very hard problem, and nowadays many difficulties and open questions are far from being solved, in spite of the efforts of a number of long-termed research groups throughout the world. Most of these problems arise when moving from the laboratory to real-world conditions. Nonetheless, ASR has resulted highly effective in a variety of applicative scenarios: dictation, which is automatic generation of written text from the speech signal, access to remote automatic services on the telephone line, and control of machines.