Research On Speech Endpoint Detection Method

Speech endpoint detection means finding the starting and ending points of an uttered word or speech segment in the presence of background noise. It plays an important role in speech signal processing which includes speech recognition, coding, and transmission and so on. An effective endpoint detection method can not only correctly identify the endpoint of speech, but also reduce the data processing time, improve efficiency, and save the data storage space.This paper introduces the speech production model and the human auditory. It analyzes the speech signal preprocessing procedures, including framing and windowing. It briefly reviews the common endpoint detection algorithm including short-time energy, zero-crossing, spectral variance, Mel-Frequency Cepstrum Coefficients and spectral entropy. And it summarizes the ideological, flow and characteristic of the algorithms and gives the graphs of the characteristic parameters used by the algorithms. It proposes two new endpoint detection algorithms.(1) An improved adaptive sub-band selection spectral variance method is proposed in this paper. Since the frequency energies of different types of noise are concentrated on different frequency bands, the bands with much noise can be discarded accurately. And we select the useful bands adaptively to yield more accurate frequency information. It improves the discriminability between speech and noise so that it becomes easier to detect endpoint and has higher accuracy. Meanwhile, it reduces the processing data and improves the performance of system. The method is applied in the Gauss continuous HMM speech recognition system. Experiment results show that the algorithm improves the accuracy and noise immunity of recognition system.(2) A speech without leading mute segment endpoint detection algorithm is proposed in this paper. The thresholds-based methods work on the assumption that the first few frames are leading mute segments. It sets thresholds according to the leading mute segments. If the value of the characteristic parameters of an unknown speech frame is larger than the threshold, the frame can be regarded as speech. Otherwise, it is regarded as noise. If the speech signal doesn't meet the assumption, the threshold predefined will be not available any more. And it lead to the failure of the endpoint detection. It avoids threshold setting problem by using FCM method. And experiment results show that it performs high performance in detecting speech without no leading mute segment.