Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Speech Modeling with State Constrained Markov Fields over Frequency Bands

V. Arsigny, Gérard Chollet, Guillaume Gravier, Marc Sigelle

ENST-TSI and CNRS-URA 820, Paris, France

In this paper, we propose a model of speech segments in the time/frequency domain. This model is based on Markov random field (MRF) modeling and is an extension of our previous work on multi-band models with MRF. In this new approach, the time model in each band is defined on the constrained state space of strictly left-right Markov chains and a non-stationary synchrony model between the frequency band is added. We derive algorithms for parameter estimation and for segment scoring. Finally the model is tested and compared to HMM on an isolated word recognition task. Results show the interest of the synchrony model for test data corrupted with additive white noise but they also point out some weaknesses of the training algorithm which must be improved.

Full Paper

Bibliographic reference.  Arsigny, V. / Chollet, Gérard / Gravier, Guillaume / Sigelle, Marc (2000): "Speech modeling with state constrained Markov fields over frequency bands", In ICSLP-2000, vol.3, 155-158.