Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

A Comparison of Three Noisy Speech Recognition Approaches

Olivier Siohan, Yifan Gong, Jean-Paul Haton

CRIN-CNRS & INRIA Lorraine, Vandoeuvre-les-Nancy, France

We compare 3 recent approaches dealing with speech recognition in noisy environment. The first approach is based on stochastic model combination of noise and speech . Given a clean speech model based on speech trajectories and an HMM noise model, this method aims at deriving a noisy speech model, in order to recognise noisy speech. In the second approach, we perform a mapping between the noisy and the clean speech space. The noisy speech is recognised after mapping to the clean space, using clean speech models. In the last approach, LDA is used as a preprocessing, and the training and testing environmental conditions are identical. On a 206 isolated word recognition task under different noisy environment, LDA gave the best results. The model combination proved to be efficient at high SNR, but performances fell down at low SNR. The mapping approach showed to be very robust, but led to the lowest recognition rate at high SNR.

Full Paper

Bibliographic reference.  Siohan, Olivier / Gong, Yifan / Haton, Jean-Paul (1994): "A comparison of three noisy speech recognition approaches", In ICSLP-1994, 1031-1034.