Supervised and Unsupervised Speaker Adaptation in Large Vocabulary Continuous Speech Recognition of Czech

摘要

This paper deals with the problem of efficient speaker adaptation in large vocabulary continuous speech recognition (LVCSR) systems. The main goal is to adapt acoustic models of speech and to increase the recognition accuracy of these systems in tasks, where only one user is expected (e.g. voice dictation) or where the speaking person can be identified automatically (e.g. broadcast news transcription). For this purpose, we propose several modifications of the well known MLLR (Maximum Likelihood Linear Regression) method and we combine them with the MAP (Maximum A Posteriori) method. The results from a series of experiments show that the error rate of our 300K-word Czech recogniser can be reduced by about 9.9 % when only 30 seconds of supervised data are used for adaptation or by about 9.6 % when unsupervised adaptation on the same data is performed.