The Algorithm Study And Application Of Speech Endpoint Detection

Speech endpoint detection(also known as Voice Activity Detection VAD) is analysis the speech signal in the noise environment, and detect the existence of the speech signal. Speech endpoint detection is usually used in speech coding, speech enhancement and speech processing systems, to reduce the speech coding rate, save communication bandwidth, reduce the energy consumption of mobile devices, improve the rate of right recognition.In speech signal analysis, First of all, it is required to judge the noise signal of input, to find out the speech segment of the signal, reduce the amount of data in processing and improve the efficiency of speech processing. The method of traditional double threshold on speech endpoint detection have high detection accuracy in pure speech. But in the actual noise environment, especially under low signal to noise ratio, the correct rate of endpoint detection is too low, this paper is based on the premise of gender information at different speaker, improved the algorithm which is wavelet energy entropy endpoint detection, improved the accuracy of endpoint detection. The main contents and results of this paper are as follows:1. This paper presents a method of speech signal analysis based on speech attribute statistics. The existing speech analysis methods focus on the short-time energy, short-time zero crossing rate, pitch, formant frequency, Mel cepstral coefficient characteristics, In this paper, according to the pronunciation characteristics of different speakers from the short term energy variance, Mel frequency range variance, MFCC and other aspects of the distance variance characteristics and other aspects of the voice signal analysis, to 239 dimensional data extracted from the speech signal, Using Relief[1] feature selection algorithm to reduce dimension, establish a reasonable set of features. Experimental results show that the recognition accuracy of speech information can be improved obviously after the introduction of speech attribute statistics.2. According to the speaker’s pronunciation characteristics, introduced into the concept of fuzzy membership function to the speech signal of speaker’s gender. Established a fuzzy membership function model based on the pitch frequency curves of different speaker language, this model can make a preliminary judgment on the gender information of the speaker. On the basis of analysis the gender fuzzy membership degree, the decision tree model is used to identify the speech documents which cannot accurately identify the gender information of the speaker. The experimental results show that the hybrid model has great improvement on the identification of gender information in the low SNR condition, the recognition effect is better.3. Under the premise of accurate identification of gender information, this paper analyzes the advantages and disadvantages of wavelet algorithm and wavelet energy entropy algorithm in the application of speech endpoint detection, and the wavelet energy entropy algorithm has improved in terms of the accuracy of operation. Finally, the improved wavelet energy entropy algorithm is used to test and analyze the noisy speech files. In different noise background, the signal to noise ratio is 5dB, the result show, the algorithm can accurately detect the speech segment. It significantly reduces the amount of information loss, improves the accuracy of endpoint detection.