Adaptation of reference speaker HMM word models for speaker-independent recognition

An algorithm for the adaptation of HMM (hidden Markov model) word models which have been trained on one or more reference speakers for the recognition of a larger population of speakers is described. A baseform transcription for each word is obtained by clustering the HMM states of the reference speakers´ models. Speaker-independent distributions for each clustered state are trained once and for all on a core set of words spoken by a large number of speakers. New words are added by training models on utterances from the reference speaker, generating a baseform transcription for the new words using the reference speakers´ clustered states, and substituting the corresponding speaker-independent distributions