Research On Broadcast News Audio Structure Analysis

Broadcast news audio structure means segmenting broadcast news audio into independent and stable audio units according to their content and analyzing audio data in them to obtain their corresponding audio scenes. Broadcast news audio structure is not only helpful for in-depth broadcast news audio information processing and analysis but also helpful to assist content-based broadcast news video analysis and retrieval. This paper mainly focuses on technologies related to broadcast news audio structure analysis, including audio feature extraction, audio segmentation and audio classificatioin. We have got the following three aspects of contributions:Firstly, in audio feature extraction aspect, UBM(Universal Background Model)-based MFCC segmental feature and two chroma-based segmental features: Intra-Frame standard variance mean and Inter-Frame standard variance mean are proposed, experimental results validates the effectiveness of our new features. Besides, Orthogonal experiment design method is introduced to choose specific features for different program recognition tasks.Secondly, in audio segmentation aspect, Believable Degree based audio segmentation algorithm is presented. Audio stream is searched using a sequential fixed-size analysis window in order to avoid accumulative errors. In the analysis window, to avoid detection errors due to threshold setting and hard threshold judging, the Believable Degree value for every hypothesized change point is computed and the true change point is detected according to the change trend of Believable Degree value. Experimental results demonstrate that our algorithm is superior to KL2-based algorithm, HMM-based algorithm, BIC-based algorithm and Entropy-based algorithm.Thirdly, in audio classification aspect, the paper puts forward SVM-GMM which combines advantages of both SVM and GMM. Considering characteristics of broadcast news audio, the"coarse to refine"strategy is employed which first coarsely classifies broadcast news audio into speech, music, speech over music and background using SVM, then recognizes the concrete program types for each audio class using GMM. Experimental results demonstrate that our algorithm outperforms cascade algorithm, NFL(Nearest Feature Line) algorithm, HMM algorithm and SVM algorithm.