HMMs and OWE neural network for continuous speech recognition
Nicolas Pican, Dominique Fohr, João Fernando Mari · 2002
The phonetic context has a large effect on stop consonants in a continuous speech signal. Therefore, recognition systems that model allophones using context-dependent hidden Markov models have been implemented (Lamel and Gauvain, 1993). HMMs have a great ability for segmentation in the temporal domain but have some difficulties in recognition because the MLE training (maximum likelihood estimation) is not discriminant, whereas discrimination is one of the abilities of artificial neural network models. In the last three years we have developed a new ANN model named OWE (Orthogonal Weight Estimator). The principle of the OWE is an ANN that classifies an input pattern according to the contextual environment. This new ANN architecture tackles the problem of context dependent behaviour training. Roughly, the principle is based on a main MLP (multilayered perceptron) in which each synaptic weight connection value is estimated by another MLP (an OWE) with respect to context representation. In this paper, we present a hierarchical system for phoneme recognition: first the system segments the input signal using 48 context independent HMMs. Then the stop consonants are reordered by an OWE ANN. Experiments on TIMIT show 78% correct recognition rate on the 6 stop consonants (/p, t, k, b, d, g).