Algorithm Study Of Vehicle’s Voiceprint Recognition Lock Based On GMM

With the increasing popularization of cars and people’s security consciousness, as well as the escalating of anti-theft products, vehicle anti-theft system is accepted by more and more users. The diversifying anti-theft methods and the improving performance of anti-theft system, are due to the development and applying of the biological recognition technology. The main task of anti-theft system is to identify the identity of user. Currently, the biological recognition technologies that used to identifying the identity of user are fingerprint recognition, face recognition, iris recognition, voiceprint recognition, signature recognition. Voiceprint recognition technology is a technology to extract the speaker’s information from the speaker’s speech and to identify the identity of speaker. In order to make the speaker recognition technology more practical, we need to overcome the influence of noise in environment. In this paper, based on the application situation of vehicle anti-theft system, which using the voiceprint recognition technology, analyze and eliminate the noise that can be got in touch with during using, and raise the recognition rate of system which surrounded with noise. The main work contains the following aspects:(1) According to the classification of noise, which voiceprint recognition lock can be exposed to while in practical applications, like periodic noise, speech noise and wideband noise, we use speech enhancement which has good effect to eliminate the noise. The speech enhancement technology can improve the anti-noise ability of recognition system’s front-end, and weaken the effect of mismatch between training environment and identifying environment on recognition rate. This article presents several de-noising algorithms:adaptive filtering method, spectral subtraction, Wiener filtering method, comb-form filtering. After comparing these algorithms, non-linear spectral subtraction and Wiener filtering method have better de-noising effect. In order to guarantee the effect of de-noising algorithm when dealing with speech with different SNR, and considering to the characteristic of embedded system and surroundings of application, we choose the improved spectral subtraction de-noising method which combining the superiority of Wiener filter and adaptive filter. The result of experiment shows the improved method can ameliorate the SNR of signal. Meanwhile, in the process of endpoint detection, utilize the improved double limits endpoint detection based on auditory masking effect, which weaken the influence of noise on the system recognition rate further.(2) Present several characteristic parameters, and compare the different combination of LPC, MFCC and the difference of them, study the effect of these parameters on improving performance of recognition system. In order to guarantee the performance of the system, we select MFCC and first-order differential MFCC, which consider the speech intra and inter frame information, and improve the steady of system.(3) United the feature of practical application and embedded platform, study the speaker recognition technology based on DTW and technology based on GMM. After comparing the different factors which effect the recognition rate, such as different number of people in testing, different length of speech, different content of text, different mix-num of Gauss, the experimental results showed that, the recognition rate can be guaranteed when the mix-num equal to16or32.(4) Build a prototype test system, which is based on Sunplus SPCE061A as the core processor, and matched with the basic data I/O interface, ADC and DAC interface, audio I/O interface, program download interface. In this platform, the test of vehicle voiceprint recognition system is completed. Based on the platform’s resource and the performance of the system, define the recognition method and the scope of system.