Excited commentator speech detection with unsupervised model adaptation for soccer highlight extraction

Soccer highlight detection is an active research topic in recent years. In this paper, we present our effort to detect an important audio keyword - excited commentator speech, which contributes to a state-of-the-art soccer highlight extraction system. We propose an approach of using statistical classifier based on Gaussian mixture models (GMMs) with unsupervised model adaptation. The excited speech and normal speech are modeled as two GMMs, and are updated to compensate for the acoustic mismatch between training and test data via Maximum a posteriori (MAP) adaptation, starting from the pre-trained GMMs. The adaptation is operated in an unsupervised mode, since the correct classification of the test data is not known, and a first pass of detection using old GMMs is performed to produce hypothesized classification results. Experimental results demonstrate the effectiveness of the proposed approach. Based on the excited speech detection alone, we can recall 87% of the goal events.