Voice Activity Detection Based On Vehicle Embedded System

With the development of automotive industry, the demand for car electronic equipment is growing faster and faster. After several years'development, the speech recognition technology now can be used in the actual situation. Building a speech recognition system on an electronic system is a trend now in car electronic industry. Voice activity detection (VAD) is an indispensable part of car electronic equipment with speech recognition system. The accuracy of the voice activity detection module has a great impact on the speed and accuracy of the speech recognition system. This paper focused on the algorithm of voice activity detection in car noise environment. A successful voice activity system can improve the speed and accuracy of the speech recognition, and make the car electronic system easy and smooth to use.In this paper, we propose a VAD method based on the cochlear model and a linear discriminant function. Noise is reduced by wiener filter, and nonlinear sub-band decomposition is carried out based on the cochlear model in frequency domain. In each sub-band, two order statistic filters (OSFs) are used to estimate the signal SNR. Finally, a linear discriminant function is used to combine sub-bands SNRs. The combined result is used for final decision. Besides, further, a new VAD algorithm is tested based on matrix eigenvector difference. One dimensional signal is converted into two dimensional signals, and image processing method is used to detect the endpoint of the voice signal. Although this algorithm fails to use on the embedded system due to large computation, the idea of converting one dimensional voice signal to two dimensional signals is important to VAD.Experiments based on 863 annotated 4 regional accent speech corpuses and car environment speech corpus has shown that the proposed algorithm maintains a high accuracy in car noise environment, and the computation of the algorithm is suitable for embedded system. The application of the algorithm to car embedded system improves the speed and accuracy of the embedded speech recognition system obviously.