The Technology Research Of Endpoint Detection And Keywords Detection Of Speech

As an important way to human-computer interaction,Speech recognition has broad application prospects. Keywords recognition is an important research field of the speech recognition, whether in flexibility or efficiency or use value speech recognition is better than continuous speech recognition. Therefore in recent years, Keywords recognition technology is a hot spot of research in the field of speech recognition.Firstly, this paper introduces the related technologies used in keywords recognition, including speech preprocessing , characteristic parameters extraction, acoustic layer HMM model, linguistics model, The keywords search and keywords verification. This paper is mainly in pretreatment stage of endpoint detection module, keyword search modules and believability confirm module.In the pretreatment stage mainly analyses the current endpoint detection methods, double threshold method and spectral entropy method is discussed in detail, and points out their advantages and disadvantages respectively. In order to improve the robustness of speech endpoint detection in low SNR environments,an improved speech endpoint detection method based on spectral subtraction and adaptive multi-band spectral entropy is proposed. The core idea of this method is: Fistly, the additive noise is removed using spectral subtraction, and background noise estimate value is updated timely; and then, improved adaptive multi-band spectral entropy is used to detect the endpoints for the enhanced speech. Experimental results indicate that this method has good detection performance. Compared with the traditional method, this method improves the endpoint detection accuracy, and good detection capability in low SNR environments.In the keyword search module, this paper has discussed on two kinds of speech identifier structure: N - Best and syllable lattice, whether in detection rates or running speed syllable lattice is better than the N - Best syllable lattice, Therefore this paper system adopts syllable lattice structure. And then improved the keyword search algorithm--dynamic programming. In order to generate the scenario hit of keywords, a token transfer algorithm is used in syllable lattice. Then in the module of confidence confirmation, the ultimately recognition results of keywords is obtained from imaginary hit of the last module. The data of the experiment shows that by using this paper detection methods,system detection rates raise approximately 10 percent. However,The false alarm rate raise one percentage points.Finally, according to the achievement of this paper, using Visual C++ development tools combined with HTK speech toolkit, a railway speech ticketing system was developed .experiment shows that the work has done is helpful to improve the performance of the recognition system.