Evaluation of RASTA approach with modified parameters for speech enhancement in communication systems

The purpose of speech enhancement techniques is to improve quality and intelligibility of speech without producing any artifact. The speech enhancement algorithms are designed to suppress additive background noise and convolutive distortion or reverberation. The need for enhancement of noisy speech in communication systems increases with the spread of mobile and cellular telephony. Calls may originate from noisy environments such as moving vehicles or crowded public gathering places. The corrupting noise is not always white rather it is colored and contains reverberation. The currently employed noise suppressors in communication systems use spectral subtraction based on short time spectral attenuation (STSA) algorithms as a preprocessor in speech coder. They can perform well in white noise condition but failed in real colored noise environments with different SNRs. This leads to the use of RelAtive SpecTrAl (RASTA) algorithm for speech enhancement which was originally designed to alleviate effects of convolutional and additive noise in automatic speech recognition (ASR). RASTA does this by band-pass filtering time trajectories of parametric representations of speech in the domain in which the disturbing noisy components are additive. This paper evaluates the performance of RASTA algorithm for white and colored noise reduction as well as suggests modifications in parameters and filtering approach to perform quite well than original RASTA approach. The NOIZEUS database is used for objective evaluation in different noise conditions with 0 to 10dB SNRs. The results shown here give improvements compared to spectral subtraction methods.