Ph.D. project: Hidden Markov Models and Neural Networks

Hidden Markov Models (HMM) are statistical models that are widely used in Automatic Speech Recognition and molecular biology. The parameters (emission and transition probabilities) of a HMM can be estimated from a set of examples by using a Maximum Likelihood (ML) training algorithm. In small voca- bulary speech recognition a HMM is trained for each word in the vocabulary, whereas in large vocabulary speech recognition subword HMM's are used (e.g. phoneme HMM's). During recognition the likelihood of each HMM is calculated, and the observed sequence is classified according to the highest likelihood. Since each model is trained using only the sequences assigned to it, it is obvious that training by ML gives non-discriminative models, i.e., the models are not trained to discriminate between words. A discriminative training method called Maximum Mutual Information (MMI) has therefore been developed and successfully applied to a range of applications.
There has recently been a widespread interest in combining neural networks and
HMM's for speech recognition. If neural networks are used to estimate probabilities in HMM's,
it is possible to estimate the weights in the neural network and the parameters
in the HMM at the same time by using a gradient descent algorithm.
The intention of this project is to analyze and develop algorithms for training combined neural network and HMM models.
Investigations have been carried out using the hybrid for recognition of five broad
phoneme classes in continuous speech (the TIMIT database). The obtained
are promising. Furthermore work has been carried on a
more real-world task (the recognition of 39 phonemes in the TIMIT database) also with promising results.
Publication.