Study On Voice Activity Detection Methods In Heavy Noise Environments

Speech endpoint detection is accurately detecting the beginning and ending of speech signal from background noise. In speech signal processing, the accurate endpoint detection can improve the speech recognition accuracy and recognition speed. At present, the accuracy of speech endpoint detection can be satisfactory in quiet environment. However,the performance of the detection severely degrades in actual different kinds of noise environments. As a result, the study on speech endpoint detection is particularly important in heavy noise environments.Firstly, the digital speech signal processing and some common speech endpoint detection methods are summarized and analysed. Some experiment results and improvements are also shown in the paper. Two new methods of speech endpoint detection in heavy nosise environments are proposed. One is based on discrete cosine transform and improved spectral entropy. The speech is enhanced with the help of the discrete cosine transform at the beginning, which effectively remove the noise of noisy signal and improve the speech endpoint detection accuracy. Then,the method of improved spectral entropy is proposed, which can more accurately detect the location of speech endpoint. Another is based on cepstral distance and short-time energy is proposed.The parameter of short-time energy is introduced, which has no relation with cepstral distance, and the planar discriminative rules in speech endpoint detection is made. Comparing spectral entropy and short-time energy with the first method, the latter not only has higher endpoint detection accuracy rate, but it also effective to detect the border between the characters of words in continuous speech, better robustness and anti-interference. The second method can be achieved easily and has better environmental adaptability, with regard to the beginning of voiced signal obviously has effective endpoint detection results. The performance of this method is better than the cepstral distance and the short-time energy. It can improve the endpoint detection accuracy in heavy noise environments, but there' s no more obviously processing. Endpoint detection is made to speech in different kinds of noise environments, evaluation and analysis are given to detection results. Experiment results show that the above two methods are efficient in different kinds of heavy noise environmentsAt the end of the paper, the two proposed methods are summarized. Questions which should be improved in the future and the perspective of endpoint detection are pointed out.