Recognition-Time Speaker Adaptation in a Tied-Mixture HMM Continuous SpeechRecognizer

摘要

All speech recognition systems, whether speaker-independent or speaker-dependent,require large amounts of training data to estimate the model parameters and, generally, the more training data available, the better the recognition performance. To improve the recognition performance of a system for a new speaker without having to train entirely new models, adapting the existing models during the recognition process is a practical solution. This report describes an investigation into the subject of recognition-time speaker adaptation of a tied-mixture HMM recognition system, with the goal of implementing a system which adapts to a new speaker during the course of its usage.