A Research Of Algorithm For Voice Endpoint Detection In Complex Noisy Environment

Accurate Speech endpoint detection in adverse environments is very important for robust speech analysis, speech synthesis and speech recognition. Although voice endpoint detection technology in quite environment has reached a high accuracy, but the performance of system will significant decrease in the practical application as the changeable noise. Ifspeech endpoint detection technology movesinto practical use, the robustness issues must be overcome. The endpoint detection technique in low SNR noiseenvironment is very important.With the arm of applied voice activity detection technology and the emphases of system robustness, this thesis deeply discusses every main aspect of isolated-word and continuous statement endpoint detection technology in noisy environment. Through the systematic research and experiment on endpoint detection problems, a complete research system of robust voice endpoint detection is formed, which includes voice database, adaptive filter algorithm and a delay segmentation strategy based on classification criteria. Based on it, a practical voice endpoint detection system can be developed. These results mentioned above can be described concretely in the follow aspect:⑴Voice endpoint detection experiment systemDepth study of voice signal mathematical models and different voice signal characteristic value and the way to extract them. TIMIT standard pure speech database and NOISEX-92 standard noise library are collected. A noise mixed voice platform which ensure the post-test repeatability is established according to metric of noise.⑵Voice enhancement algorithmFor conventional adaptive filtering algorithms on the convergence rate and stable accuracy and computational complexity of irreconcilable introduced European search algorithm, the algorithm made several improvements to reduce the calculation accuracy, which greatly improved the convergence speed and stability, through Comparative experiments verify its performance close RLS algorithm, and its calculation are much smaller. SNR in MOS and evaluation methods, also received a higher performance.⑶Endpoint detection algorithmDetailed analysis of the commonly used two-threshold endpoint detection method, endpoint detection algorithm based on spectral entropy and endpoint detection algorithm based on fractal theory. The introduction of the arrangement of entropy, a non-linear kinetic parameters as well showing non-linear characteristics of the speech signal. Proposed a delay segmentation strategy: to be able to determine the characteristic frequency ratio parameter crude endpoints and using permutation entropy difference algorithm on this basis to determine the precise endpoint to endpoint as a starting point the exact division of speech signal, the speech signal obtained fragments in accordance with the classification criteria to eliminate noise signal errors caused segmentation.⑷System implementationUse Matlab GUI tool to achieve whole endpoint detection system interface, using the second chapter of speech endpoint detection database to compare and test different methods. Experiments show that the method proposed in this paper has better detection results than conventional two-threshold-based spectral entropy approach, especially in the case of low SNR, the basic method to achieve the effect based on fractal. But with the effect of the text after the filtering effect of the program than any other method. And because of the simple arrangement entropy algorithm easy to implement, real-time performance of the algorithm is very good, the complexity of its calculation is much smaller than the fractal method.