Research On Nonlinear Filtering Speech Endpoint Detection Algorithm Rased On Microphone Array

With the development of society, the concept of the artificial intelligence is now known by more and more people. The topic of artificial intelligence is a comprehensive topic, where the speech recognition is very important part. Only when the machine can recognize what people said, it is proven as smart.In speech recognition, the voice endpoint detection technology is paid more and more attention, because whether the endpoint detection is correct or not directly affects the efficiency and accuracy of the back-end voice recognition. In the speech signal endpoint detection, the recognition rate of the traditional detection method will be declined in the low signal to noise ratio environment. In this thesis, adaptive linear filtering voice endpoint detection mode based on microphone array is studied to improve the overall algorithm performance. The main research contents include as follows:(1) Aiming at the problem that the recognition rate in the traditional endpoint detection algorithm is declined in the case of low signal-to-noise ratio of speech signal, the linear filter speech endpoint detection structure based on microphone array is studied. Through the improved algorithm of speech enhancement based on microphone array, the initial cancellation of speech noise is realized, which makes the correct rate of the whole endpoint detection be increased by nearly 10%.(2) For many background noise is short-term stable or unstable situation in voice conversation environment, the Legendre nonlinear expansion function in the structure of the filter is studied, which can make the speech enhancement algorithm better remove the non-linear component in the background noise. Also, it enhance the adaptive performance and robustness of the endpoint detection algorithm.(3) Aiming at the case that the misjudgment or missed detection occurs by short-term zero-crossing rate and short-term energy value under the influence of background non-coherent noise, the voice endpoint detection algorithm using parameter combination between Teager energy value and the short-time zero-crossing rate based on EMD decomposition are studied. By comparing endpoint detection accuracy under different background noise environment, it is proven that the proposed algorithm is better than the traditional algorithm, and has better robustness and anti-noise performance.(4) Through the sound recordings recorded in the silence room of Nanjing University of Information Science and Technology, and combined with the background noise in the NOISEX-92 noise library, the actual environment is simulated and the adaptability of the algorithm is verified. At the same time, the performance of the proposed algorithm is judged by subjective evaluation of PESQ score and objective evaluation of SNR.