The Study Of Speech Endpoint Detection Methods Based On Cepstrum Characteristic And Voicing Features

Endpoint detection serves as the core technology in the preprocessor of the speech recognition system, its directly determines the succeed of the whole speech recognition system to a great extent. As we know, the object of speech signal processing is effective speech signal in facts, so it is essential for us to find the begin and end point of the effective speech signal in input signal. Endpoint detection aims at finding the begin and end point of the speech signal in input signal, we could also call that dynamic speech detection.In this paper, it first presented bring-model of the speech signal in briefly, including the characteristic of speech signal and noise signal, some theory about buffer window, separate frame processing of speech signal and so on. Second, it introduced some endpoint detection methods proposed by the domestic and international scholars in recent years, and analysed about detection methods with simulation results given in particular. Afterwards, it introduced a estimate method which can manifest periods of the speech voicing features. Finally, the ameliorative speech endpoint detection algorithm adopted in this article is proposed, including:One, endpoint detection base on cepstrum characteristic (cepstrum distance) is executed with the speech signal given, and save the information into cache memory.Two, the methods of detecting the about the speech signal are adopted, including detecting the speech harmonic information of the speech signal base on speech matrix chart, the voicing features base on the speech periods estimated and the method of detection voice and unvoice base on the wavelet transform.After analysis and simulation experiment, we can find that the endpoint detection method proposed in this paper can preferably detect the endpoint of speech signal, and have better veracity, outstanding noiseproof feature, strong robustness.