Adaptive recognition method based on posterior use of distribution pattern of output probabilities

摘要

Proposes a new adaptation scheme for speaker-independent recognition. The basic idea lies in the change of the likelihood from ordinary HMM scores to combined observational scores. The new likelihood is computed based on a combination of HMM scores which we call a 'pattern of output probabilities distribution' (POPD). The system needs to calculate only the POPD for each new speaker. Re-estimation of acoustic model parameters is unnecessary. Preliminary experiments on Chinese isolated-syllable recognition indicate the method's effectiveness.