Research On Train Whistle Recognition Method Based On Wavelet MFCC And HMM

With the rapid development of society, how to extract useful information in massive junk information has become a hot topic, meanwhile the research and application in sound recognition has also attract more and more attention. Traditional sound recognition methods can’t meet the increasingly high demand now, especially in the train whistle recognition. Train whistle recognition includes two parts:the extraction of sound characteristic parameter and the classification of sounds. Using this technology enables automatic recognition of the train whistle, in order to promote the lookout of the train crew. Train whistle belongs to abnormal sound, currently abnormal sound recognition methods follow the traditional methods of speech signal processing, vocalization of speech signal is fixed, and the energy is relatively stable, while in train whistle recognition, sounding mechanism of each sound is quite different, the instantaneous energy is large, and identifying whistle is easily affected by background noise. Therefore traditional methods of speech signal processing are not well adapted to train whistle recognition.Addressing the above questions, this thesis studied how to construct train whistle recognition algorithm based on the speech signal feature Mel Frequency Cepstrum Coefficients and Hidden Markov Model. Firstly, this thesis made an brief introduction to the extraction method of MFCC and difference MFCC. By analyzing the train whistle characteristics during the recognition process, proposed WMFCC feature extraction method on the basis of Wavelet Transform. After comparing various types of classifiers, this thesis selected Hidden Markov Model as a classifier used in train whistle recognition. Finally, according to the proposed train whistle recognition algorithm based on wavelet MFCC and HMM, recorded the train whistle on-site and simulated the recognition in MATLAB.Results of experimental studies showed that using the method based on wavelet MFCC and HMM in the train whistle is feasible. Meanwhile, the algorithm’s extending to special sounds recognition proved that it is superior than traditional methods in explosions, gunfire, and door slamming recognition.