Using spatial cues for meeting speech segmentation

This work investigates the validity and accuracy of using spatial cues with Time-Delay Estimation (TDE) as a method of segmenting multichannel recorded speech by speaker location. In environments such as meetings where speakers do not significantly alter position, segmentation by speaker location essentially leads to segmentation by speaker 'turn'. The proposed system calculates location information using TDEs and spatial cues extracted from multichannel meeting audio recordings. This location information is then input into a simple segmentation algorithm. Experiments have been performed on both theoretical and real meeting recordings with non-overlapping speakers, and theoretical recordings with overlapping speakers. Segmentation results reveal the most robust cue to be a combination of spatial information and TDEs. This cue combination leads to greater segmentation accuracy for classifying individual speakers and detecting overlapping sections than using spatial cues or time-delay information alone.