Research On Driver's Speech Enhancement Method In Car Environment

Speech recognition system in car environment is designed for the purpose to liberate the driver's hands, provide driving pleasure and promote travel safety. In recent years, along with the surge in China's automobile usage, speech recognition system in car has get widespread attention both at home and abroad. Some related products have come out since 2000. Existing voice recognition products can reach a recognition rate about 97% in a quiet environment, but in the car environment, it is difficult to achieve a practical standard. The main reason is that, highway condition under car environment is very complicated and all kinds of environmental noise will affect the recognition rate. To make things worse, it will affect driving safety.An important part of speech recognition system in car is to distinguish the driver's speech first. Which is the key of speech recognition of drivers. Therefore, this thesis focuses on the driver's speech enhancement problem under the car environment. The noise of car environment includes voice and non-voice noise, and the traditional enhancement methods cannot de-noise the voice interference by non-drivers. Therefore, this thesis considers enhancing the driver's speech by speech separation. The work of this thesis includes the following two aspects:(1) In this thesis, the author designed a method based on the combination of wavelet threshold de-noising method and blind source separation. Speech separation problem for drivers under the car environment belongs to underdetermined blind source separation, which is difficult for implement, and blind source separation algorithm cannot improve the signal to noise ratio effectively. So this thesis makes the wavelet threshold de-noising method as a pretreatment. This thesis first uses the pretreatment to de-noise the noisy speech for filtering out non-voice noise, then it can get a mixed voice which contains the driver and many other people's speech. Thus, underdetermined blind source separation can change to a non-underdetermined blind source separation problem, and it can greatly improve the signal to noise ratio. At last, the thesis takes fast independent component analysis method to separate the mixed voice. As a result, the thesis can get the independent speech of each speaker.(2) To make use of the voiceprint features of speech to select driver's voice from many people's voice. The author choose the Mel frequency cepstrum coefficient(MFCC) as the voiceprint features, and the vector quantization(VQ) as the method of a pattern matching. During the training phase, the thesis trains the MFCC of driver as a pattern of VQ. During the testing phase, extracts the MFCC of each speaker's speech obtained from the blind source separation, then calculates the distance to the pattern of driver and selects the minimum one as the driver's speech.The simulation results of the method show that the proposed method in this thesis for driver's speech enhancement can reach a certain de-noising effect to non-voice noise under the complex road noise environment, and it can improve the signal to noise ratio. After removing the non-voice noise, the complicated underdetermined blind source problem will change to a simple non-underdetermined blind source problem. In addition, blind source separation can clearly separate the mixed speech into each people's independent speech component. With voiceprint matching, the method can accurately identify the driver's speech component without the support of microphone array. This method is simple and easy to implement on the hardware platform under car environment.