This project investigates the problem of labelling segments in a speaker-tracking system. A mathematical representation of each segment is sought which encaptures the speaker-dependent information available. It is shown that both the covariance matrix and the Maximum Likelihood Linear Regression (MLLR) matrix provide such a representation with over 90% success rate in a speaker identification task.
Several alternative distance metrics to measure the ``closeness'' of the segments are investigated and it is found the covariance data performs best on those based on the mathematical means of the eigenvalues of one matrix relative to another. The MLLR matrix is by contrast found to work best with elementwise metrics confirming the hypothesis that the individual elements of the matrix are more significant in this case.
Several hierarchical clustering schemes are then investigated and shown to produce speaker-specific groups on two and three speaker problems. A full-scale implementation is then described and tested on data from the 1996 Broadcast News database. A new criterion for evaluating the clusters is defined and shown to be a good indication of speaker split. Six clustering schemes are evaluated using this new criterion and the discriminative Lance-Williams scheme is found to perform the best. Furthest neighbour clustering is also shown to perform well in some cases.
Tree diagrams for the best cases of three and four clusters are presented and explained in terms of the clustering strategy which produced them. They illustrate the feasibility of such a system, with only a few segments being obviously mis-classified. Possible improvements for the system are then discussed and finally recommendations for further work are given.
If you have difficulty viewing files that end
which are gzip compressed, then you may be able to find
tools to uncompress them at the gzip
If you have difficulty viewing files that are in PostScript, (ending
'.ps.gz'), then you may be able to
find tools to view them at
We have attempted to provide automatically generated PDF copies of documents for which only PostScript versions have previously been available. These are clearly marked in the database - due to the nature of the automatic conversion process, they are likely to be badly aliased when viewed at default resolution on screen by acroread.