Large vocabulary hidden markov model based speech recognition

Abstract

AbstractThis paper gives an introduction to the use of Hidden Markov Models and Information Theory principles for speech recognition. The use of these techniques has led to powerful large vocabulary speech recognition systems developed in the recent years. The use of Hidden Markov Models for stochastic modeling of speech and the basic modules of a Hidden Markov Model based speech recognition system are explained. It is shown how the speech waveform can be transformed into acoustic features and into prototypes which characterize the acoustic state space of the given speaker. The phonemes of the language can be represented by Hidden Markov Models. The statistical parameters of these Markov models can be obtained during a training session. During recognition the input to the linguistic decoder is the acoustic label string resulting from the speech signal of a spoken sentence. The decoder picks the sentence which maximizes the probability of the word sequence of this sentence if the given label stream is seen. The paper concludes with a more detailed explanation of the acoustic processor of the speech recognition system and some proposed algorithms for its improvement which can lead to a better system performance.