This work adresses the problem of adapting to noise Hidden Markov Models Speech Recognition Systems used in car environment The recognizer is generally trained in clean conditions, while recognition is performed in noisy conditions. Three techniques are presented to adapt recognizers to new environments: speech enhancement using nonlinear spectral subtraction, transforming references using a mean linear transformation (learned by linear regression) and adjusting the mean vectors of HMM states using the knowledge of the ambient noise. Original results with our database gave about 79 % (90 km/h) and 72 % (130 km/h) of recognition rate. The three techniques proposed show improvements in terms of recognition scores between 18% and 21% (90 km/h) and between 22% and 26% (130 km/h). Therefore, these methods are very promising and even suggest several further developments.
Cite as: Mokbel, C., Barber, L., Chollet, G. (1992) Adapting a HMM speech recognizer to noisy environments. Proc. ETRW on Speech Processing in Adverse Conditions, 211-214
@inproceedings{mokbel92_spac,
author={C. Mokbel and L. Barber and Gérard Chollet},
title={{Adapting a HMM speech recognizer to noisy environments}},
year=1992,
booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
pages={211--214}
}