As with everything else in life, there is more that could be done, and it could not be otherwise in this thesis.
One topic within the meeting domain processing that has received quite some attention recently is speaker overlap detection. As such, it refers to the detection of the segments where more than one speaker is talking at the same time, and the output of an appropriate ID for each participant. In NIST RT06s evaluation the main metric included overlap for the first time and multiple research labs (including the author) researched techniques for its detection without any success in reducing the overall error. There is still work to be done in detecting when more than one person is speaking, which should come from the beamforming module (where speakers are well determined by their location) and from the diarization module (where data with multiple overlapping speakers has special acoustic properties compared to single speakers). Also in ASR systems for meetings it would be very beneficiary to create multiple signals in overlap regions, each one derived from the steering to every speaker.
Another area where there research should be directed is the creation of strong links between ASR transcription output and diarization. Although it is well established the use of diarization algorithms to help the ASR systems in model adaptation, it has only been briefly studied the use of ASR for diarization. It could be useful in a number of areas, like the definition of possible speaker change-points (ranging from boundaries at word level to discourse level), or in the assignment of speaker ID's (or the correct name) based on the transcription content (in which LIMSI has done some research). Also, both areas could benefit by the combination of plausible speaker labels with ASR N-best words for each instant, it being useful at decoding level to reduce the errors in both the ASR and diarization tasks.
In the topic of discourse modeling, speaker diarization could benefit from research on ways to model the turn-taking between the speakers. By using information at a higher level than pure acoustics, the transition probabilities between speakers could be appropriately set to help the decoding. One of such possible high level information is easily noticeable in broadcasted news where an anchor speaker is very probable to speak after every other speaker. Similar analysis could be made in the meetings domain and possibly classify the meetings into several types (more fine grained than the current lecture/conference classification) to apply different models to. Possible types could me: moderated meetings (with one person acting as anchor), structured meetings (people following an order, without anchor) and unstructured (where everyone intervenes at random, supposedly with higher amount of overlap regions). Also, it could be considered to split the meeting into several parts with each person's participation dependent on which part/topic the meeting is in.
One of the objectives of this thesis was to increase the robustness of the system to mismatches between the development data and the test data, and to make the system parameters less sensitive to the processed data by obtaining parameters more linked to the acoustics and by eliminating any model training step from the system. It has been shown in the experiments section that an important step forward has been taken in that direction. There is still more that can be done in this topic towards eliminating as many tuning parameters as possible, letting the algorithms select such parameters solely from the data. Also important is to better understand the underlying processes that lead the system to score very differently depending on every particular meeting (leading to ``easy'' and ``difficult'' meetings).
Finally, although current systems in the RT evaluations are defined with no application in mind, trying to be adaptable to any possible application, this poses a burden in the capacity of such systems to obtain the optimum score and makes them more computationally intensive as most of the algorithms used for diarization are iterative. It would be interesting to explore particular areas of application where, for example, the number of speakers in a meeting are known. This particular information would probably change the way that speaker diarization algorithms are designed and would allow for lower DER scores, most probably in the region where speaker identification techniques are nowadays.