We present an approach for recognising continuous speech in the presence of an additive noise, based on model adaptation. The method consists in transforming the parameters of acoustic models to reduce the acoustic mismatch between a test utterance and a set of clean speech models. We assume that speech is modelled by a set of Stochastic Trajectory Models (STM). The mean vectors of STMs are adapted using a set of linear transformations. The transformations are derived from a small labelled adaptation corpus, so that the likelihood of the adaptation corpus given the adapted models is maximised. Experiments performed on different additive noises and for various signal-to-noise ratio (SNR) show that the adaptation scheme significantly increases the accuracy. For SNR from 12dB to 36dB, we observed that the performance of the linear regression is similar or better than the performance obtained when training and testing the recogniser in noise.
Bibliographic reference. Siohan, Olivier / Gong, Yifan / Haton, Jean-Paul (1995): "Noise adaptation using linear regression for continuous noisy speech recognition", In EUROSPEECH-1995, 465-468.