Research On Unsupervised Speaker Clustering And Its Implementation

With the popularity of the Internet and multimedia acquisition device, people can getmore audio data that has an explosive growth. Managing audio data through manualannotation is too expensive and subjective, and can’t meet the need for massive audio datamarking. So content-based audio retrieval and search technology came into being.Speaker Clustering is an important part of the content-based audio index and retrieval. Itsmain task is labeling the divided voice segments by speakers, means that the audio data fromthe same speaker is labeled as the same and the labels are differed from the speakers. Thispaper focus on Unsupervised Speaker Clustering and the spectral clustering algorithm basedon feature affinity matrix and model affinity matrix, and designed a system for speakerdiarization and clustering.The main research work is described as follows:1. Analyze basic principles and methods of several speakers clustering algorithmcommonly used, focus on the basic principles, processes and key issues (building affinitymatrix in spectral clustering and the automatically estimation of the number of speakers).2. Test the operating efficiency of the spectral clustering algorithm based on featureaffinity matrix. Achieve a spectral clustering algorithm based on model affinity matrix by anadaptive Gaussian mixture model. The experiment proved that the algorithm based on modelaffinity matrix greatly reduced the running time and had good performance that can be similarto the algorithm based on feature similarity matrix.3. Design three experiment programs to evaluate the performance of the algorithm,respectively using CCTV News, conference speech and telephone speech as the experimentaldata, and evaluate the algorithm by two evaluation criteria. At last, compare and analyze theresults.4. Design the unsupervised speaker diarization and clustering system, it is divided into fivemodules: Speech Reading Module, Endpoint Detection Module, Feature Extraction Module,Speaker Segmentation Module and Speaker Clustering Module. And the Speaker ClusteringModule provides two spectral clustering methods for users to select. The system is built byMFC and allows users setting their own parameters, so that it has the visualization feature and simple operability.