The Endpoint Detection Algorithm Of Speech Based On Fractal Dimension

The endpoint detection technology of speech signal is to accurately determine starting point and ending point from a section of speech signal. Thus it can distinguish speech and non-speech signal. Effective endpoint detection can not only reduce the amount of data collection and save the processing time, but also can eliminate interference from the silent and the noise. It can improve property of speech recognition system. Besides it can reduce bit rate of the noise and the silent in speech coding so improve the coding efficiency. Therefore endpoint detection is very important in speech processing.It is a bit difficulty to detect endpoint accurately in low SNR, especially in silent segment and pre-and post pronunciation. This paper summarized the typical endpoint detection algorithm, including the algorithm based on short-time energy and zero-crossing rate, the algorithm based on LPC cepstrum, the algorithm based on entropy function, the algorithm based on HMM and the algorithm based on sub-band average energy variance. The paper analyzed the different feature and presented the part of the simulation results. Those algorithms can have a good performance when it is quiet or has a small noise. But the result has a rapid decline when the environment is bad and SNR is low. The paper proposed three methods of endpoint detecting in noise environment. The first is the endpoint detection based on fractal dimension. It utilizes fractal dimension superiority and overcomes the difficulty of decision threshold in noise environment. The second is the endpoint detection based on fractal dimension and fuzzy RBF neural network. This method combines the advantages of both fractal dimension and information entropy and neural network which avoid threshold setting. The simulation result shows that this method is better in accuracy of endpoint detection in low SNR. The third one is endpoint detection based on 1/f fractal signal wavelet model and fuzzy RBF neural network. The experiment shows that it has a better effect in normal noise environment. The algorithm is easy and adaptable to environment.