Dsp-based Voice Activity Detection And Neural Network Adaptive Filter

The main task of the speech signal detector is to judge the signal or not, and then send the results to the main board to control transmission or receiving state.In this paper, in order to detect voice segment and silent segment, analyzing theory basic about speech detecting, discussing different characters between voice and noise, studying signal spectral entropy and AMDF's effect on speech signal, an algorithm that combined spectral entropy with AMDF is put forward, and adaptive neural network filtering is used, SNR is improved, speech is enhanced. Then voice detector's design is accomplished based upon the DSP chips.When studying algorithm, we divide the interference into white noise and high distortion monosyllabic wave, square wave, and analyze and deal with the signal in frequency domain and time domain separately. In frequency domain, making use of FFT and the maximal discrete entropy theory, analyzing the short-term speech spectrum, we construct a kind of entropic function to detect voice segment and silent segment. In time domain, we take out the short-time signal with rectangle window, and calculate the short-time AMDF of the signal.The result shows that the algorithm is good in high SNR. But the SNR is poor in vehicle communication. So adaptive neural network filter is designed before signal processed. The simulation shows the algorithm have high correct probability.In hardware design, we take full advantage of the multichannel buffered serial ports(McBSP) and the direct memory channels of the DSP chip. Thus the import and export of the digital signal do not impropriate the CPU time. Also the circuit of this speech signal detector is simple and the size of PCB is small enough.