Speaker Adaptation Technology And Its Key Words In The Telephone Channel Detection System Applications

The research of keyword spotting in speech recognition has made great progress in recent years and it has been applied in the telephone channel. However in the factual application, the laboratory "success" spotting system can not satisfy the practical requirement in robustness, brainwave and adaptation ability. This point can be seen in the speaker independent speech recognition system. The reason of degrade is the influence brought by variational speaker and environment. When there is a mismatch between the testing and the training conditions or the presence of low-level background noise, the error rate will increase. This paper analyzes the robustness of interferential factor to speaker and studies how to increase the adaptive ability of system.This paper discusses speaker adaptation technique from speaker normalization, model parameter adaptation and speaker clustering. It first introduces the most widely used model-based adaptation methods: MAP adaptation and MLLR adaptation, Eigenvoice adaptation and SMAP adaptation are then studies based on them. These methods have their own strongpoint and shortcoming and can be used in diversified condition. This paper combines these methods successful in keyword spottinf system. Experimental result indicates for the new speaker, the average recognition error rate dropps 6.3% using 5 adaptation sentences; 32.6% using 30 adaptation sentences. Different from traditional speaker clustering, this paper introduces coordinate axis-based speaker clustering which can make sure the new speaker's sort using few corpora. Speaker normalization includes the Cepstrum Mean Normalization and Vocal Tract Length Normalization. During the implementation of keyword spotting system, this paper developes the robust module of speaker adaptation through uniting three method of speaker adaptation above. Furthermore, the system applies noise reduction technique: a method combining speech enhancement and compensation which can achieve approving result especially in the case of low signal-noise rate. At last, this paper gives the conclusion and the research direction in the next period.