Abstract:

Voice Endpoint Detection is challenging, especially in nonstationary noise and a low signal-to-noise ratio(SNR), so this paper proposes a novel Robust Voice Endpoint Detection method fusing Burg spectrum estimate and long-term signal variability(LTSV). This method uses a novel long-term signal variability measure, by which the degree of nonstationarity in various signals can be indicated. Comparison with the traditional Voice Endpoint Detection method based on signal features, this method's detection performance has been greatly improved under the condition of a low signal-to-noise ratio and nonstationary noise. Also, Burg spectrum estimate is proposed, which improves the LTSV parameter discrimination degree, thus further improving the detection accuracy. Simulation results show that in comparison with the standard Voice Endpoint Detection method, the new method's accuracy is generally improved by more than about 6%, which shows that the new method has better robustness in the non-stationary noise and low signal-to-noise ratio environment.

Key words: voice endpoint detection, long-term signal variability measure, Burg spectrum estimate, low signal-to-noise ratio, nonstationarity