Endpoint Detection Technology Research

Speech endpoint detection is a key problem in many speech applications, such as speech analysis, speech synthesis and speech recognition. At present, the accuracy of speech endpoint detection can be satisfactory in quiet circumstance, but with the noise polluting and the circumstance changing, its performance will degrade severely. Only the robust problem is solved, the technique of speech endpoint detection can be taken to application. So endpoint detection at low SNR is crucial for good speech recognition accuracy.We summarize most of present methods of speech endpoint detection. In result, a novel approach that finds robust features for endpoint detection in a noisy environment is proposed. In this proposed method, we applied KC computation complexity and C₀ computation complexity into the speech endpoint detection. These features can be used to distinguish speech/noise in low SNR. It is a creative thought and trial.In the experiment, noisy speech with different SNR is analyzed by conventional speech endpoint detection method compared with the algorithms proposed in the paper, including energy, MFCC cepstrum distance, spectral entropy, KC computation complexity and Co computation complexity, which are programmed with the tool of Microsoft Visual C++ 6.0. We have three speech databases (YOHO English continuous speech database, one Chinese continuous speech database and one isolated speech database). This paper adopts COLEA, is MATLAB software for speech analysis, to add noise from NOISEX 92 into speech data with certain SNR. The proposed algorithm is shown to be well suited for the detection of speech endpoint and is very robust for different types of noise, especially for low SNR. The visualization of experiment results are realized by MATLAB 6.5. Experimental results indicate that C₀ computation complexity method has extremely robust and high accuracy, meeting the requirements of robust speech recognition system. The experiment results show these feature are valid, and have broad application prospects.