Analysis on MAP and MLLR based speaker adaptation techniques in speech recognition

摘要

Speech recognition system produces a text output corresponding to the given speech input. A speaker-dependent (SD) recognition system results in a higher recognition performance when compared to a speaker-independent (SI) system. Speaker adaptation techniques like maximum aposteriori (MAP) and maximum likelihood linear regression (MLLR) are applied to an SI system, in order to get a recognition performance similar to that of an SD system, with minimal amount of data. The main focus of this paper is to analyse the performance of the adaptation techniques, applied to the recognition system for different amount of adaptation data. In this work, a speech recognition system is developed using Tamil speech corpus. Cross-gender speaker adaptation is performed by varying the adaptation data. It is observed that when the adaptation data is very minimum, around 30s, the recognition performance of MLLR adapted system results in 45.76% when MAP adapted system resulted in 42.44%. When the adaptation data is increased to 5min, the overall recognition performance is improved by 6% for MAP adaptation over MLLR adapted recognition system.