A Simulated-Data Adaptation Technique for Robust Speech Recognition

This paper proposes an efficient acoustic model adaptation method based on the use of simulated-data in maximum likelihood linear regression (MLLR) adaptation for robust speech recognition. Online MLLR adaptation is an unsupervised process which requires an input speech with phone labels transcribed automatically. Instead of using only the input signal in adaptation, our proposed simulated data method increases the size of adaptation data by adding noise portions extracted from the input speech to a set of pre-recorded clean speech, whose correct transcriptions are known. Various configurations of the proposed method are explored. Evaluations are performed with both additive and real noisy speech. The experimental results show that the proposed system achieves higher recognition rate than the system using only the input speech in adaptation and the system using a multi-conditioned acoustic model.