Speaker Diarization Module Experiments

This section analyzes each of the proposed improvements to the mono-channel diarization module proposed in this thesis. The algorithms that are analyzed in this section are:

• Number of initial clusters (section 4.2.2)

• Cluster complexity selection (section 4.2.2)

• Cross-Validation EM training (section 3.3.3)

• Friends-and-enemies clusters initialization (section 4.2.1)

• Frame purification (section 4.3.1)

• Segment purification (section 4.3.2)

• Multiple feature streams automatic weighting (section 5.3.1)

All experiments use the same baseline system as described in 6.1.1. This uses the new hybrid speech/non-speech detector and the RT06s beamforming system when necessary. In order to evaluate all possible diarization working conditions, three different tasks are run:

• The SDM system uses only the single SDM channel as defined by NIST for the RT datasets used in the experiments.

• The MDM mono-stream system uses the beamforming of the multiple available channels to obtain a single acoustic channel for use by the system. The same configuration is used for the diarization of this channel as for the SDM channel.

• The MDM multi-stream system uses the beamformed signal as well as the extracted TDOA values to add information to the diarization on the location of the speakers. Although this system uses the same acoustic channels as the pure MDM task, this evaluation condition is added as it behaves differently to the data than any of the other tasks.

Experiments were performed in two steps to find the optimum parameters for all algorithms by using the development set. First, each algorithm was tested alone against the proposed baseline. One or two parameters for each algorithm were searched for the optimum configuration. Then, the different algorithms in the order of individual improvement (from most to least) were used together and the optimum parameters were searched for again, obtaining the optimum system according to the development. Finally, the evaluation set was used to test how the system performs with unseen data.

The optimization of the parameters is done using the arithmetic average of the three considered systems, with the exception of the multi-stream weight computation algorithm which only affects the system using the TDOA feature stream.