Model-based compensation of the additive noise for continuous speech recognition. Experiments using the AURORA II database and tasks

摘要: In this paper we apply a model-based compensation method to cancel the effect of additive noise in Automatic Speech Recognition systems. The is formulated statistical framework order perform optimal given observed noisy speech, model describing statistics speech recorded clean reference environment and estimation recognition environment. estimated using first frames sentence be recognized frame-by-frame algorithm performed, so that procedure does not constrain real-time systems compatible with emerging technologies based on distributed recognition. We have performed experiments under conditions AURORA II database for tasks developed as standard reference. Experiments been carried out including both, multicondition training approaches. experimental results show improvements performance when proposed applied.