Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Dynamic HMM Selection for Continuous Speech Recognition

T. Hain, P. C. Woodland

Cambridge University Engineering Department, Cambridge, UK

In this paper we propose a dynamic model selection technique based on hidden model sequences (HMS). HMSmodelling assumes, that not only the actual state sequence is unknown, but also the model sequence given aparticular sentence. This allows more than one model tobe used for a particular phone in a certain context. Themost appropriate model is determined locally rather thana priori globally by the acoustic probability of that modeltogether with a probability that this model is producedin a particular phone (or model) context. Experimentson the Resource Management corpus show significant im-provements in word error rate over phonetically model - and state - tied triphone hidden Markov models (HMMs).Initial results on the Switchboard corpus also show im-provements on a much more dificult task.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Hain, T. / Woodland, P. C. (1999): "Dynamic HMM selection for continuous speech recognition", In EUROSPEECH'99, 1327-1330.