Baseline Systems

Taking as a reference the blocks diagram in figure 3.5, experiments were conducted on three of the main blocks, namely the filter&sum module, the speech/non-speech module and the mono-channel speaker diarization module. For each block a baseline was defined to suit its characteristics and to allow for the development of its optimum parameters selection. The initial Wiener filtering of the signal was not analyzed as it was used without modification from its original implementation outside of the scope of this thesis.

The baseline system used for the experiments in the diarization module and in the speech/non speech module corresponds to a modified version of the broadcast news system presented for the NIST RT04f evaluation as described in section 3.1. This corresponds to a mono-channel system (or Single Distant Microphone, SDM, in the meetings domain) with the following main differences from RT04f:

• The speech/non-speech (spnsp) detector used in the experiments for the hybrid spnsp algorithm is composed of a two-states HMM model trained with meetings data, as it was used in the RT05s evaluation and explained in section 3.1. For the speaker diarization module the proposed hybrid spnsp detector was used instead, with parameters equal to the values used for the RT06s evaluation (see Anguera, Wooters and Pardo (2006b), (Anguera, Wooters and Pardo, 2006a)). These use the parameter values optimized in the spnsp experiments section.

• During the agglomerative clustering processing the same speaker turn minimum duration is applied as in the broadcast news system (3 seconds). Before the output of the resulting segmentation, a final segmentation step is performed using the same speaker models but reducing the minimum duration to 1.5 seconds to allow for smaller speaker turns to be properly detected.

• The HMM acoustic models used in the segmentation of the data do not have any maximum time constraint, as explained in section 4.2.3, to allow the speaker segments to be as long as the acoustics dictate. As shown in Anguera, Wooters and Hernando (2006a) it does not change much the DER of the systems but allows for longer speaker segments to be created.

• A few bug fixes regarding floating point values inexactitudes were resolved which slightly changed the system outputs.

• The BIC-based stopping criterion is used in all experiments in order to stop clustering when the optimum number of clusters is reached.

The baseline system used for experiments on the beamforming module is composed of the submission to RT06s NIST evaluation campaign. This contains all the modules as explained in section 5.2 and their parameters optimized using a subset of 10 meetings from the development data available for RT06s.