Technology Of Audio Signal Processing Based On Wavelet Neural Network

With the rapid development of computer technology, intelligent devices are increasingly favored. During the process of intelligent devices design, Human-Machine Interface (HMI) technology has been highlighted gradually. Language, the unique trait that separates human beings from the others, became the chief media in HMI. Thus, the research on digital audio signal based on human languages has become a domain of modern computer technology that of great importance. However, the diversity and the low anti-jamming ability of digital audio signal made it difficult to achieve ideal precision of recognition.Research in this paper started from the complexity of Chinese audio signal. Firstly, wavelet threshold de-noise algorithm is implemented to solve the problem of signal de-noise. Also a new wavelet threshold de-noise algorithm based on wavelet coefficient energy and zero rate was proposed in this paper, which resulted in a wonderful effect in the de-noise experiment of low SNR audio signal, and achieved better precision in feature extraction and recognition. Next step is the pre-processing of audio. After frame separation, short-time energy and zero rate of each frame would be calculated and then the boundary of de-noised audio signal was selected using the classic dual-threshold. Feature extraction was carried out with the universally used LPCC and MFCC as feature parameters. And audio signals of different length were warped into a designated number of frames using dynamic time warping (DTW) technology, which were considered the feature parameters. Wavelet technology was discussed thoroughly from following aspects:the set of knot number, selection of wavelet base function, correlation of input and output, and learning strategy. In this process, a wavelet network dynamic learning algorithm based on self-adaptive learning rate was proposed in this paper, which could implement the leaning based on decreasing grads for network parameters while at the same time approach the dynamic modification of learning rate. The above algorithm not only overcome the partial minimum problem of classic neural network, but could avoid the slow convergence as well, and finally led the wavelet network to a static and fast constrain effect. Ultimately the algorithms mentioned above were applied into the experiment of recognition of Chinese audio signal.The result indicated that for detection and recognition of Chinese audio signal, wavelet neural network holds a shorter training time and higher recognition precision. The preliminary de-noise process is helpful to gain a better effect on audio feature extraction, and the anti-jamming ability of the audio signal recognition system has been increased. This paper is significative to the design of intelligent recognition system for audio signal.