Speaker Detection in Broadcast Speech Databases

Abstract

Experiments have been carried out to assess the feasibility of detecting target speaker segments in multi-speaker broadcast databases. The experiemental database consists of NBC Nightly News broadcasts. The target speaker is the news anchor, Tom Brokaw. Gaussian mixture models are constructed from labelled training data for the target speaker as well as background models for other speakers, commercials, and music. Four labelled 30-min. broadcasts are used for testing. Mel-frequency cepstral features, augmented by delta cepstral features are calculated over 20 msec. windows shifted every 10 msec. through a broadcast. Likelihood ratio scores are calculated for each test frame averaged over blocks of frames with a specified duration. The block scores are input to a detection routine which returns estimates of target segments boundaries. The range of best results obtained over the test broadcasts is 82% to 100% detection of target segments with segment frame accuracy ranging from 86% to 95%. 0 to 2 false alarm segments are detected over each 30 min. broadcast.