| Speech segmentation mechanism plays an important role in many applications,the final result of which can be further utilized to study multimedia information retrieval,speaker clustering,speaker tracking.For example,when we combine speech segmentation mechanism with speaker clustering,the Speaker Diarization system can be achieved,which can provide rich information related to speakers effectively.Being a part of Speaker Diarization,Speech segmentation mechanism can solve two problems simultaneously,which include speech activity detection and speaker change point detection.Due to the diversity of data type,the insufficiency of modeling data and the shortage of prior knowledge,there is certain difficulty in the implementation of the algorithm.For speech activity detection,support vector machine is utilized to classify the speech and non-speech segments,and time-frequency feature is employed to discriminate different non-speech types.In speaker change points detection,a combination of prosodic feature and Bayesian Information Criterion is used to validate the change points,thus improving the accuracy and stability of the system.Speech segmentation mechanism mainly consists of feature extraction,speech activity detection,speaker change point detection module.The key points of this paper encompass the following aspects:(1)Implement and analysis the system of content analysis for speech segmentation mechanism,evaluate the experiment parameters,find out the potential problems and propose solutions.(2)In speech activity detection,compare different classification algorithm of speech and non-speech classification,choose the best algorithm;study different feature representation and classification algorithm,implement enhanced non-speech classification.(3)Compare the advantages and disadvantages of two baseline systems for speaker change point detection.Design a new prosody based system,propose false alarm compensation proposal,and improve the accuracy and stability of the system. |
PDF Full Text Request