HMM INITIALIZATION AND TRAINING FOR A ROMANIAN LANGUAGE CONTINUOUS SPEECH RECOGNIZER

摘要

This paper is about training of acoustical models in continuous speech recognition framework. It is known that the performance of an automatic speech recognition (ASR) system depends on how accurate it models the speech. In modeling phone-based HMM, the training stage is critical for the system accuracv. We have focused on HMM parameter initialization and re-estimation of these parameters by embedded training. There are two methods presented herein for initialization: the first one is based on segmented training data while the other one makes all the models identical in the so called flat start scheme. After models initialization, some aspects about embedded training are considered. Finally, experimental results with two gender dependent systems are tested with the two methods. We concluded that hand-labeled data is not necessary for training HMM, flat start being a more practical solution to model initialization.