The MAP-SPACE denoising algorithm for noise robust speech recognition

Abstract

We present a new and simple algorithm (MAP-SPACE) for robust speech recognition which can be seen as an hybrid approach between a denoising and an adaptation technique. This algorithm first models clean and noisy training speech using GMMs and then build a denoiser which depends only on the GMMs parameters. Given observations in a new environment, the noisy speech GMM is adapted and the parameters of the adapted GMM are then used in the denoiser to compute clean feature estimates. The MAP-SPACE algorithm requires in principle relatively few adaptation data, does not require transcription and does not make any assumption on the corrupting noise. We report preliminary experiments on the Aurora2 database. The results show that MAP-SPACE achieves very good performances, sometimes approaching those of the matched models, in both SNR and noise type mismatch conditions