Study Of Real-Time Acoustic Detection Based On Semi-Supervised Learning

With the development of modern information technology, especially network and multimedia technology, multimedia data has become a major part of the data transmission on the internet. Audio is an important form of multimedia. Owing to the rapid growth of the amount of audio data, it becomes more and more important to quickly and effectively retrieve the needed audio information from large-scale database. The method of machine learning can reduce the human intervention and more intelligently process the audio data, while traditional audio retrieval technology requires abundant labor force. According to the user's specific requirements, machine learning methods can find useful knowledge from mass data by model building. In this dissertation, the semi-supervised learning strategy is employed for audio retrieving and a method to detect sound from the specific environment is proposed.Firstly, the MFCC character of the low frequency coefficients after three-layer wavelet decomposition, the frequency spectrum centroid of primary data and the spectral entropy possessing good anti-noise performance were extracted from audio data by means of muti-resolution characteristics of wavelet transform. And then 14-D feature vector was constructed by the mean of MFCC, the variance of centroid and spectral entropy.Then, a new Tri-training algorithm named AR-Tri-training base on the assistant strategy was proposed in order to resolve two problems:introducing noise while the Tri-training was learning and low efficiency of unlabeled examples. Based on the few labeled samples, AR-Tri-training algorithm could eliminate noises in learning process and make full use of the data by taking advantage of rich information strategy and the assisted learning strategy.Finally, the audio learner was applied to special scene and a real-time detection algorithm was presented to distinguish the audio type existed in scene. In this algorithm, audio data was sampled by 2s, then the features of real-time audio data were extracted and inputted to the audio learner. The test results and temporal distribution were outputted after audio learner's analyzing. The experimental results show that AR-Tri-training can't only remove the noise in the learning process, but also take full advantage of unlabeled samples. The recognition rate of the proposed algorithm is obviously improved compared with Tri-training. In addition, the corresponding detection system shows that the performance of detection system is better than the system based on Tri-training.