Continuous speech recognition with a robust connectionist/markovian hybrid model

This paper introduces a novel combination of Artificial Neural Networks (ANNs) and Hidden Markov Models (HMMs) for Automatic Speech Recognition (ASR), relying on ANN non-parametric estimation of the emission probabilities of an underlying HMM. A gradient-ascent global training technique aimed at maximizing the likelihood (ML) of acoustic observations given the model is presented. A maximum a-posteriori variant of the algorithm is also proposed as a viable solution to the "divergence problem" that may arise in the ML setup. A 46.34% relative word error rate reduction with respect to standard HMMs was obtained in a speaker-independent, continuous ASP, task with a small vocabulary.