Continuous Speech Recognition Using Dynamic Bayesian Networks: A Fast Decoding Algorithm
Résumé
State-of-the-art automatic speech recognition systems are based on probabilistic modelling of the speech signal using Hidden Markov Models (HMMs). Recent work has focused on the use of dynamic Bayesian networks (DBNs) framework to construct new acoustic models to overcome the limitations of HMM based systems. In this line of research we proposed a methodology to learn the conditional independence assertions of acoustic models based on structural learning of DBNs. In previous work, we evaluated this approach for simple isolated and connected digit recognition tasks. In this paper we evaluate our approach for a more complex task: continuous phoneme recognition. For this purpose, we propose a new decoding algorithm based on dynamic programming. The proposed algorithm decreases the computational complexity of decoding and hence enables the application of the approach to complex speech recognition tasks