A comparison of noise reduction techniques for robust speech recognition

This report presents the integration of several noise reduction methods into the front-end for speech recognition developed at IDIAP. The chosen methods are : Spectral Subtraction, Cepstral Mean Subtraction and Blind Equalization. These different methods are studied from a theoretical point of view, their implementation is described and are tested on the Numbers95 speech database. A good noise robustness is obtained by combining two of these methods, like Spectral Subtraction with Cepstral Mean Subtraction or Spectral Subtraction with Blind Equalization. The later combination is found to be more appropriate for real recognition systems since it is frame synchronous. A comparison with Jah-RASTA-PLP is also given.