Location and Extraction of Broadcast in News Video Based on QGMM and BIC

摘要

An algorithm on location and extraction of broadcast in news video is proposed in this paper. Firstly, input audio stream is divided into speech and non- speech segments by VQ (Vector Quantification) after a set of new features representing audio segments' time-variant characteristics are extracted, including HZCRR (High Zero-crossing Rate Ratio), LSTER (Low Short-time Energy Ratio) and HBFERR (High Basic-frequency-energy Rate Ratio), etc. Then a QGMM (Quasi Gaussian Mixture Model) is presented to describe the speaker's identity and BIC (Bayesian Information Criterion) is used to detect speaker change. Finally speaker clustering is carried out with BIC, and location and extraction of broadcast is realized based on rules. Satisfactory results from experiments prove the effectiveness of this algorithm.