HMM Speech Recognition Technology Based On Keywords

Voice keyword recognition is an important research direction of speech processing. with the40years’continuous development, it is used in various social and economic applications in lives, while it is used relatively small in the military speech. on the one hand, the research of National defense voice is the lack of a professional research team, on the other hand, its particularity led to the complexity of the research,thus these factors determine the development and application of low-level. Based on the reasons above, starting from the study of the particularity of the military voice message to investigate and study them, We mainly talk about the following aspects:1. Speech de-noising, introduced some common methods of de-noising and its advantages and disadvantages.Based on them,with the shortcomings of different effect of different types of noise of the traditional de-noising algorithm,we base on the wavelet transform,focusing on the threshold de-noising method based on wavelet transform and improve the threshold function.The experiments results show that compared to other de-noising methods,this way can significantly improve the military voice-noise ratio,and it is more in line with actual needs.2. Based on de-noising, in order to better detect the start point of the voice,we discuss zero law endpoint detection algorithm,.In order to handle the blurred results of detection,we take the difference of the forward and reverse of the voice to determine the endpoint,and the experiments results show that this detection method is more accurate.3. Feature extraction, we research the extraction methods of linear characteristic parameters and discrete wavelet transform features,and based on these,we study the contributions of differential Mel cepstral feature extraction method based on discrete wavelet transform of improvement in system robustness.4. Identify links, We study three mainly issues:hidden Markov model assessment model decoding and model training, and we take appropriate preventive measures to handle data underflow in model assessment and the decoding process. In aspect of keywords confirming,we research common confirm methods based on acoustic model,the dynamic ranking information and the posterior probability of the three confidence.and we discuss the advantages and disadvantages of these three methods by experimental test, finally established system of confidence confirmation in this article.5. The implementation of the recognition system, based on the experimental results, make full use of the algorithm of this paper, the design of military voice keyword recognition system, including software interface, voice library.