Hybrid recognizers combining hidden Markov models and multilayer perceptron

This paper describes some approaches for hybrid recognizers combining hidden Markov models (HMM) and multilayer perceptrons (MLP). One of these approaches employs MLP as a post-processor for HMM while the other uses HMM to segment the speech signal for MLP. The performance of hybrid recognizers is compared with discrete HMM and multilayer perceptrons. All of the implemented recognizers were speaker-independent and a 50-word vocabulary spoken in Brazilian Portuguese was employed in their evaluation. The speech signal was parametrized using mel-frequency cepstrum coefficients, mel-frequency cepstrum coefficients with cepstral mean removal, energy and delta coefficients