Reverberant speech enhancement using linear prediction residual signal

This thesis proposes a new method to improve perceptual quality of reverberant speech by giving more emphasis to the segments of the speech signal having higher direct path signal component. It is based upon analysis of short segments (2 ms) of data to enhance regions in the speech signal having high signal to reverberant component ratio (SRR). Using short segment analysis, the method identifies and manipulates the linear prediction residual signal in three different regions of the speech signal, namely high SRR region, low SRR region and only the reverberation component region. An overall weight function is derived to modify the linear prediction residual signal. The weighted residual signal is then used as an input; along with the linear prediction coefficients; to a fifth order all pole filter to obtain perceptually enhanced speech. The performance of the proposed algorithm is illustrated using spectrograms and subjective evaluations.