CMAC spectral subtraction for speech enhancement

One of the major problems in speech signal enhancement and cancellation of additive noise is the availability of a reference signal. A comprehensive and efficient technique for speech enhancement based an extension of the spectral subtraction method is developed. In our proposed model, enhancement is achieved by using a class of associative memory based on the cerebellar model arithmetic computer (CMAC) as a robust method to estimate the reference signal. CMAC can learn very fast and it can approximate a wide variety of nonlinear functions. Thus the learning algorithm of CMAC can be integrated with the spectral subtraction method to produce a system that allows the noise estimate to be learned adaptively. The effectiveness of the architecture is demonstrated on speech corrupted with very low signal-to-noise ratio (from -5db to -20db) in a vehicular environment.