On-line speaker diarization aims to detect "who is speaking now" in a given audio stream. The majority of proposed on-line speaker diarization systems has focused on less challenging domains, such as broadcast news and plenary speeches, characterised by long speaker turns and low spontaneity.
The first contribution of this thesis is the development of a completely unsupervised adaptive on-line diarization system for challenging and highly spontaneous meeting data. Due to the obtained high diarization error rates, a semi-supervised approach to on-line diarization, whereby speaker models are seeded with a modest amount of manually labelled data and adapted by an efficient incremental maximum a-posteriori adaptation (MAP) procedure, is proposed. Obtained error rates may be low enough to support practical applications.
The second part of the thesis addresses instead the problem of phone normalisation when dealing with short-duration speaker modelling. First, Phone Adaptive Training (PAT), a recently proposed technique, is assessed and optimised at the speaker modelling level and in the context of automatic speaker verification (ASV) and then is further developed towards a completely unsupervised system using automatically generated acoustic class transcriptions, whose number is controlled by regression tree analysis. PAT delivers significant improvements in the performance of a state-of-the-art iVector ASV system even when accurate phonetic transcriptions are not available.
Finally, a first attempt at combining PAT and semi-supervised on-line diarization confirms the potential of PAT in improving real-time speaker modelling motivating further research in this particular direction.