Research And Implementation On Noisy Speech Endpoint Detection Algorithm

In a complete speech recognition system, the final realization and effect is not only limited to the recognition algorithm, the impact of a speech recognition system has many factors.The object of speech recognition is a speech signal, the purpose of endpoint detection in complex environments is to distinguish speech signal and non-speech signals, determining the signal’s start and end. Speech signals usually have some background sound, and the speech signal is inevitable to be affected by various noises in a real environment, the first crucial technology of speech recognition processing is speech endpoint detection. Only correctly the speech signal do endpoint detection, we can improve the accuracy of speech recognition.The research of this paper is based on the existing algorithm of speech endpoint detection. The main research focus on the following two points:This paper focus on the study of MATLAB simulation for a variety of voice endpoint detection algorithm, and on the basis of spectral entropy algorithm, proposed an improved weighting factor. Using the spectral entropy advantages and characteristics, experimental results demonstrate the effectiveness and robustness of spectral entropy algorithm. To use spectral entropy below for endpoint detection provides a very suitable algorithm.The second emphasis of this paper is the endpoint detection algorithm based on speech enhancement, and focus on the LMS algorithm and spectral subtraction. On the basis of the basic spectral subtraction, put forward the improved spectral subtraction. In solving complex problems of noise, the paper introduced the adaptive noise smoothing. Combined the improved spectral subtraction with adaptive noise smoothing, this paper does endpoint detection for noisy speech signal. Simulation results show that the improved algorithm is better able to adapt the complex noise environments, and improve the speech signal endpoint detection rate.