Lecture Room Systems

In the lecture room environment the submission consisted on primary systems for the tasks MDM, SDM and MSLA, and contrastive systems for MDM (two systems), SDM and MSLA (two systems).

Following is a brief description for each of these systems and their motivation:

• MDM, SDM and MSLA primary condition (MDM/SDM/MSLA_p-omnione): It was observed in the development data that on many occasions it was possible to obtain the best performance by just guessing one speaker for the whole duration of the lecture. This is particularly true when the meeting excerpt consists only of the lecturer speaking, but is often also achieved in the question-and-answer section since many of the excerpts in the development data consisted of very short questions followed by long answers by the lecturer. They were therefore presented as the primary submissions, serving also as a baseline score for the lecture room environment. Contrary to what was observed in the development data, the contrastive (``real'') systems outperformed the primary (``guess one speaker'') submissions on the evaluation data. Depending on what data is to be processed (the length of the lecturer turn and the amount of silence in the recordings) it might not be feasible to improve upon a ``dummy'' system with the current state of the art diarization systems.

• MDM using speech/non-speech detection (mdm_c-spnspone): This differs from the primary submission only on the use of the speech/non-speech (spnsp) detector to eliminate the areas of non-speech. On the development data it was observed that non-speech regions were only labelled (in the hand-made references) when there was a change of speakers, which never happened for the ``all lecturing'' sections. In a real system though it is important to detect these silences and not attribute them to speakers. This submission is meant to complement the previous one by trying to improve performance where between-speech silences are marked.

• MDM using only the TableTop microphone (mdm_c-ttoppur): From the available five microphones in the lecture room, one microphone (labelled as ``TableTop'' microphone) is clearly of much better quality than all the others (which can be found via an SNR comparison among the channels). It is located in a different part of the room and is of a different kind, which could be the reason for its better performance. In the evaluation data it was found by using an SNR estimator and the standard diarization is used on it. No spnsp detection was used in this system.

• SDM using the SDM channel with a minimum duration of 12 seconds for each cluster (sdm_c-pur12s): This uses the clustering system on the SDM channel. It didn't use the spnsp detector either. It was observed that using a minimum duration of 12 seconds, the issue of silences marked as speech in the reference files could be bypassed, and force the system to end with fewer clusters.

• MSLA with standard filter&sum (msla_c-nwsdpur12s): In order to combine the various available speaker-localization arrays, we used the filter&sum processing, using a random channel from one of the arrays as the reference channel. The enhanced channel obtained was then clustered using the 12 second minimum duration system.

• MSLA with weighted filter&sum (msla_c-wsdpur12s): In the time between the conference room and lecture room submissions, experiments were performed with a first version of the weighted filter&sum algorithm as presented in this thesis. It was applied to the MSLA channels in this system.