Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Maximum Likelihood Sequential Adaptation

Beth Logan

Compaq Computer Corporation, Cambridge Research Laboratory, Cambridge, MA, USA

We develop a new sequential adaptation technique for HMMs based on an incremental variant of the EM algorithm. The approach has little impact on the speed of normal Viterbi decoding and in the case of mean adaptation only, is equivalent to incremental MAP adaptation for a certain choice of priors. We apply the technique to the ARPA HUB4 broadcast news task. Here since the acoustic conditions change frequently, it is advisable to `reset' the adaptation process periodically. However, for this task, the acoustic conditions change so rapidly that it is difficult to obtain enough information for adaptation between model resets. Many existing adaptation schemes tackle this problem of data sparsity by cleverly updating unseen mixture components. We investigate an orthogonal strategy in which a set of models, each representing a different acoustic condition, is maintained and adapted. We show that small improvements in performance are possible using this approach.

Keywords: sequential adaptation, online adaptation, speech recognition, HMM.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Logan, Beth (1999): "Maximum likelihood sequential adaptation", In EUROSPEECH'99, 17-20.