HMM-Based Keyword Spotting System

As the research of speech recognition going further into application, the technology of detecting and verifying some special words in a sentence or a paragraph that be given has drawn more and more expert's attention--which is called technology of keyword recognition(KWR)or keyword spotting(KWS). This technology can be used in much applications flexibly, such as the real-time search of audio data in Internet, making operator assisted call in words by customer, automatic monitor for governmental and military affairs, etc. In all aspects of this technology the keyword recognition in continuous speech is most concerned.KWS use DTW, HMM or ANN to modeling the speech signal, the same as other speech recognition. HMM is more popular and important in KWS. In this thesis, we focus research on the keyword spotting system which use HMM. There several keystone in my work:1. Put forward a new ameliorated Viterbi search algorithms——DifferentialState Dwell Distribution-Based Viterbi search algorithm2. Confidence measures. In the keyword verification, according to making different speech features contain complementary speech signal information, give a new confidence measure which based on different feature secondary recognition. And give the best feature combination choice.3. Establish a KWS system. According to the system test result, this KWS have good spotting capability. In the process of coding this system, using Modularization Idea makes this system have good expanded potential; using reduced arithmetic makes this system have good Real-time performance.4. Experimentation. In the examination, carry on three experiments: speech featrue, search algorithms, cofidence measures. Confirm that Differential State Dwell Distribution-Based Viterbi search algorithms have better correct rate and robust performance; different feature-based secondary recognition confidence measures have good reject capability.