A comparative study of model-based adaptation techniques for a compact speech recognizer

Abstract

Many techniques for speaker adaptation have been successfully applied to automatic speech recognition. This paper compares the performance of several adaptation methods with respect to their memory need and processing demand. For adaptation of a compact acoustic model with 4k densities, eigenvoices and structural MAP (SMAP) are investigated next to the well-known techniques of MAP (maximum a posteriori) and MLLR (maximum likelihood linear regression) adaptation. Experimental results are reported for unsupervised on-line adaptation on different amounts of adaptation data ranging from 4 to 500 words per speaker. The results show that for small amounts of adaptation data it might be more efficient to employ a larger baseline acoustic model without adaptation. Eigenvoices achieve the lowest word error rates of all adaptation techniques but SMAP presents a good compromise between memory requirement and accuracy.