Performance Analysis of LP Residual and Correlation Coefficients based Speech Seperation Front End

Performance of modern automatic speech recognition systems can be enhanced either at the front end by, speech separation system or by a set of efficient features or by a powerful classifier at the backend. A speech separation front end is built using average zero crossing rate, log energy, correlation coefficient of LP residual for separating the speech frames from AWGN samples. The experimental results on TIMIT database demonstrates that the proposed speech separation system performs well over the AWGN of various SNR levels.