Hidden mode HMM using Bayesian network for modeling speaking rate fluctuation
Authors:
- 7 September 2004
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
Abstract
One of the most important issues in spontaneous speech recognition is how to cope with the degradation of recog- nition accuracy due to speaking rate fluctuation within an utterance. This paper proposes an acoustic model for adjust- ingmixtureweightsandtransitionprobabilitiesoftheHMM foreachframeaccordingtothelocalspeakingrate. Thepro- posedmodelisimplementedalongwithvariantsandconven- tional models using the Bayesian network framework. The proposed model has a hidden variable representing varia- tion of the "mode" of the speaking rate and its value con- trols the parameters of the underlying HMM. Model train- ing and maximum probability assignment of the variables are conducted using the EM/GEM and inference algorithms for Bayesian networks. Utterances from meetings and lec- tures are used for evaluation where Bayesian network-based acoustic models are used to rescore the utterance hypothe- sesobtainedfromafirst-passN-bestlist. Intheexperiments, the proposed model shows consistently higher performance than conventional models.Keywords
This publication has 10 references indexed in Scilit:
- Speech recognition of spontaneous, noisy speech using auxiliary information in Bayesian networks2003
- Time adjustable mixture weights for speaking rate fluctuation2003
- Estimating the speaking rate by vowel detection2002
- Combining multiple estimators of speaking rate2002
- Indicator variable dependent output probability modelling via continuous posterior functions2002
- Multiple-regression hidden Markov model2002
- The meeting project at ICSI2001
- Buried Markov models for speech recognition1999
- Extraction and representation rhythmic components of spontaneous speech1997
- Speech recognition using on-line estimation of speaking rate1997