This contribution addresses the problem of speaker-independent continuous speech recognition in the presence of uncorrelated additive noise. Most experiments reported thus far dealt with small vocabulary isolated word recognition. We describe how the acoustic feature extraction by an auditory model can be improved by introducing noise magnitude subtraction inside the auditory model. A computationally attractive algorithm is presented and evaluated. It clearly outperforms the standard power spectral subtraction postprocessing technique applied on the outputs of the original auditory model. The experiments show that no retraining of the speech models and no tuning of the proposed algorithm to specific noise conditions is required.
Bibliographic reference. Vereecken, Halewijn / Martens, Jean-Pierre (1995): "Recognition of noisy speech using an auditory model", In EUROSPEECH-1995, 1995-1998.