On the use of higher-order statistics for robust endpoint detection of speech

Third order statistics of speech signals are not identically zero, as it would be expected based on the linear model for voice. This is due to quadratic harmonic coupling produced in the vocal tract. Based on this observation, third order cumulants are employed to address the endpoint detection problem in low SNR level recordings due to their immunity to (colored) additive non-skewed noise. The proposed method uses the maximum singular value of an appropriately formed cumulant matrix to distinguish between voiced parts of the speech signal, and silence (noise). Adaptive implementations are also proposed, making this method computationally attractive. Results of batch and adaptive forms are presented for real and simulated data.<>