Speech processing for statina

Abstract

Reverberation reduces the speech intelligibility and degrades the performance of various speech controlled systems. A dereverberation algorithm aims to reduce the multipath effects and retrieve the estimate of the clean speech signal. In order to suit a practical environment, a blind dereverberation technique considering noisy environment with slowly varying acoustic channels is evaluated and performance is analyzed. The speech dereverberation model uses the estimated impulse response to dereverberate the speech and to improve the signal-to-noise ratio without having a priori information about the Acoustic Impulse Responses (AIRs), statistical properties of speech or the location and source, and microphones. The received noisy signals are processed by an eigenfilter block, which enhances the power of the speech as compared to that of noise. A frequency domain constraint is incorporated in the design of the eigenfilter which improves the quality of the signal, resists spectral nulls and improves Signal to Noise Ratio (SNR). A zero-forcing algorithm which eliminates the distortion caused by the AIRs and eigenfilters is used to dereverberate the speech signal. The block-adaptive implementation of Zero Forcing Equalizer (ZFE) makes the algorithm suitable to work in a time-varying condition. A study on the various dereverberation techniques is done and the blind speech reverberation technique explained above is evaluated and analyzed. The simulation results show superior performance.