Effective Speech Endpoint Detection Algorithm For Voiceprint Recognition

Speech voiceprint recognition with noise in complex real phone channel environment is still a critical challenge even the recognition method works well enough in non-noise situation. Background noise, especially dial tone of voice, which is the voice from surrounding disturbs the accuracy of recognition. One key problem of voiceprint processing is how to locate when the voice start and stop, and another one is how to remove all kinds of noise effectively. In this paper, we tackle these two problems and propose an endpoint detection algorithm which based on a double threshold method by processing short-time energy and linear prediction cepstrum distance. By compensating the high frequency part of the speech signal and the frequency spectrum of the signal become flat, we avoid the energy losing of small voice signal and improve the accuracy of detection. Our algorithm remains the principle of speech signal with little cost. Experiment shows the effectiveness of our algorithm both in public voiceprint dataset and real public security case dataset.