A Comparative Study On Endpoint Detection Methods Of Erroneous Words In Errors

The speech recognition is attracted widespread attention by domestic and foreign scholars, and has developed rapidly in recent years. Its ultimate goal is to realize man-machine communication freely. Many research achievements have emerged. However, we have to face one of the key issues is that how to separate the voice signal in practice environment effectively-namely the environment with noise. Endpoint detection technology is the key to the final recognition rate in the speech recognition process. Appropriate endpoint detection method will improve the final recognition accuracy and efficiency greatly. Moreover, there are about more than eight million people using the Yi language in China. Therefore, the study of endpoint detection for Yi language under noisy environment speech recognition has important application value and practical significance.This paper describes the research status at home and abroad of endpoint detection in the speech recognition process, and the basic theory of this technology, such as voice signal pre-emphasis, framing and windowing, etc. Then the main characteristic parameters of endpoint detection is discussed in time domain and frequency domain respectively. Moreover, several classical endpoint detection methods are analyzed in time domain and frequency domain, such as short-time energy and short-term zero rate method and short-time correlation method in time domain, frequency band variance method and spectral entropy method in the frequency domain. And noisy Chinese speech signal are performed experiments with these methods, the experimental results and analysis are given. However, for Yi speech signal under the conditions of low SNR, the accuracy of the results of classical endpoint detection algorithm is far less than our requirements. Therefore, this paper introduces the empirical mode decomposition (EMD) "Teager" algorithm, and combines with an improved zero-crossing rate method. For the Yi speech of actual recorded under low SNR environment, propose endpoint detection method based on combined improved zero-rate and EMD "Teager" algorithm. Then for this method and the classical endpoint detection methods compared and studied. Experimental results show that, the proposed method compared to the classical endpoint detection methods, can get more accurate endpoint detection results for Yi speech signal under relatively low SNR environment.