The Study Of Speaker Indexing In Multi-Channel Environment

The main content of this thesis is the study of multi-channel speaker indexing in open environment, with the purpose to build an audio engine for speaker indexing, which can search the World Wide Web for a specific speaker. Speaker indexing can be viewed as an application of automatic speaker recognition. In dealing with the variety and openness of the World Wide Web and its great differences in audio environment, this thesis tries to differentiate the audio environments and weaken its effect and the interference of channels so as to improve the performance of indexing.The first task of this thesis is to build a datum speaker indexing engine through the study about all aspects of speaker recognition and indexing, which is used for performing speaker indexing in steady environments and raising the recalling rate on the bases of ensuring the general correctness.The second one is to develop the speaker indexing technology on the bases of dynamic decision mechanism. As speaker recognition and indexing are both performed through accepting or refusing a tested audio frequency, a score threshold is to be established to process audio frequency test. Those audio frequencies with higher scores than the threshold are accepted and those with lower scores are refused. Furthermore, a dynamic generative threshold is introduced to set up an index approach of decision mechanism.The third one is the study of recognition algorithm based on channel clustering approach. It is very difficult to deal with the complicate audio frequency documents in open environment, including noise strength, distortion caused by different equipments and the loss of document compact. In this thesis, a clustering approach is introduced to solve the problem, i.e., the unknown audio frequency documents are classified according to environments and channels. The environments and channels in the same audio frequency are approximate and easier to process, and the related channel information can be drawn through the statistic analysis of all the clustering.The fourth one is the introduction of channel compensation based on clustering algorithm. The above-mentioned channel clustering algorithm supplies the necessary condition for channel compensation. Further improvement is made for the channel compensation algorithm on the base of channel clustering method, with the result of optimizing the compensation performance.Finally, this thesis brings forward several new kinds of algorithm with easy application and high performance for speaker indexing in authentic environment, which supply the base to establish a multi-channel speaker indexing system.This work is supported by National Natural Science Foundation of P.R.China(60273059), Zhejiang Provincial Natural Science Foundation for Young Scientist of P.R.China (RC01058), Zhejiang Provincial Natural Science Foundation (M603229) and National Doctoral Subject Foundation (20020335025).