Unsupervised speaker segmentation of multi-speaker speech data

A method for segmenting speech data into speaker segments by speaker, the method comprising:

• scanning input speech data with a windowed generalized likelihood ratio (GLR) function to obtain speech segments, wherein the input speech data includes a plurality of speakers;
• clustering the speech segments into one or more clusters, wherein each cluster is associated with a single speaker;
• if more clusters exist than speakers, then:
• checking overlap between segments in each cluster;
• pooling clusters that have overlap between at least one segment in each pooled cluster;
• and
• resegmenting and remodeling the pooled clusters;
• creating models for each cluster;
• and
• rescanning the input speech data with the models to resegment the speech data and obtain speech segments for each speaker included in the speech data.

A method for segmenting speech data by speaker, the method comprising:

• obtaining initial estimated segments of the speech data, wherein the estimated segments are unlabeled;
• clustering the initial estimated segments until the initial estimated segments are grouped into a final group;
• selecting one or more clusters from the final group, wherein one or more clusters corresponds to groups of estimated segments;
• iteratively modeling and resegmenting each group of estimated segments until changes in segment boundaries for the estimated segments in each group of estimated segments from a particular iteration to a next iteration are below a threshold;
• if at least one segment in a first group of estimated segments overlaps with at least one segment in a second group of estimated segments, then pooling the first group and the second group of estimated segments;
• and
• iteratively modeling and resegmenting the pooled first and second group and any remaining groups of estimated segments to generate new groups of estimated segments until the overlap between the new groups of estimated segments is below a particular level;
• and
• associating each group of estimated segments and each new group of estimated segments, if any, with a different speaker.