| In recent years, hidden Markov models (HMM's) have evolved to become the most popular technique for the application of speech recognition. However, individual HMM's have a weak discriminating power such that input observations might be mis-represented by the wrong HMM for isolated-word recognition. Thus, the main objective in this research is to find a way to enhance the discriminating power of HMM's, which in return will also lead to the reduction of errors in continuous speech recognition.;In order to enhance the performance of HMM, this dissertation attempts to apply a secondary training process from neural networks to adjust the HMM parameters. A supervised learning technique, namely learning vector quantization (LVQ), is used for the additional training procedure. However, there are two major difficulties that lie in front of the joint training of HMM and LVQ. First, LVQ has to process the input observation in real time with the knowledge of the input pattern. Unfortunately, input signals like speech are usually composed of a sequence of observations whose state sequence is not determined until the whole input sequence has entered the HMM. Secondly, LVQ is memoryless and is not able to preserve the temporal information within the speech signals, which is crucial to signals like speech.;To solve the above two difficulties impeding the joint application of HMM and LVQ, a sequential learning vector quantizer (SLVQ) is proposed in this research. During the training process, each HMM is regarded as a sequential LVQ codebook which can preserve the temporal information of speech signals. SLVQ allows the real-time processing of input speech which travels through each HMM. Experiments of speech recognition for both the isolated-word and the connected-word cases using the SLVQ training procedure have shown improvement of correct recognition. |
