Speech Recognition Under Noise Conditions: Compensation Methods
Ángel de la Torre, C. Jose, Carmen Benı́tez, Javier Ramirez Luz Garcia, Antonio J. · 2007
In this chapter, we have presented an overview of methods for noise robust speech recognition and a detailed description of the mechanism degrading the performance of speech recognizers working under noise conditions. Performance is degraded because of the mismatch between training and recognition and also because of the information loss associated to the randomness of the noise. In the group of compensation methods, we have described the VTS approach (as a representative model-based noise compensation method) and histogram equalization (a nonlinear non-model-based method). We have described the differences and advantages of each one, finding that more accurate compensation can be achieved with model-based methods, while non-model-based ones can deal with noise without a description of the distortion mechanism. The best results are achieved when both methods are combined.