Speech endpoint detection algorithm with low signal-to-noise based on improved conventional spectral entropy
Yi Zhang, Kejia Wang, Bo Yan · 2016
Endpoint detection is a typical classification problem, which means the signal will be divided into two types of speech and background noise. Under the environment with low SNR, in order to better enhance the discrimination of noise and improve the accuracy of speech endpoint detection system, this paper proposes a new type of voice parameter — sub-band energy-entropy-ratio by combining the sub-band energy on the basis of analyzing the conventional sub-band spectral entropy. Which means the ratio of short-time sub-band energy and sub-band spectrum entropy as an important parameter of endpoint detection. Experiments show the algorithm is not only fast and efficient, with strong robustness, but has strong anti-noise ability under lower SNR and can accurately detect the voice endpoint.