Integration of multi-layer perceptron and Markov models for automatic speech recognition

The paper presents the implementation of a new speech recognition system based on the integration of three semi-independent blocks: the acoustic processor (AP), which converts the speech signal into a set of robust acoustic features: the multi-layer perceptron (MLP) that maps the acoustic feature sequences to phonemes, discriminating the spectral variation from the real phonetic information; and the hidden Markov model (HMM), which produces a final identification of the entire utterance as consequence of the computations of the probabilistic phonetic observations output by the MLP.<
>