Voice activity detection (VAD) algorithms are essential for many speech processing applications, such as speaker diarization, automatic speech recognition, speech enhancement, and speech coding. With a good VAD algorithm, non-speech segments can be excluded to improve the performance and computation of these applications. In this paper, we propose a polynomial eigenvalue decompositionbased target-speaker VAD algorithm to detect unseen target speakers in the presence of competing talkers. The proposed approach uses frame-based processing to compute the syndrome energy, used for testing the presence or absence of a target speaker. The proposed approach is consistently among the best in F1 and balanced accuracy scores over the investigated range of signal to interference ratio (SIR) from -10 dB to 20 dB.

Evaluation of speaker segmentation and diarization normally makes use of forgiveness collars around ground truth speaker segment boundaries such that estimated speaker segment boundaries with such collars are considered completely correct. This paper shows that the popular recent approach of removing forgiveness collars from speaker diarization evaluation tools can unfairly penalize speaker diarization systems that correctly estimate speaker segment boundaries. The uncertainty in identifying the start and/or end of a particular phoneme means that the ground truth segmentation is not perfectly accurate, and even trained human listeners are unable to identify phoneme boundaries with full consistency. This research analyses the phoneme dependence of this uncertainty, and shows that it depends on (i) whether the phoneme being detected is at the start or end of an utterance and (ii) what the phoneme is, so that the use of a uniform forgiveness collar is inadequate. This analysis is expected to point the way towards more indicative and repeatable assessment of the performance of speaker diarization systems.

Human-based speaker diarization experiments were carried out on a five-minute extract of a typical AMI corpus meeting to see how much variance there is in human reviews based on hearing only and to compare with state-ofthe-art diarization systems on the same extract. There are three distinct experiments: (a) one with no prior information; (b) one with the ground truth speech activity detection (GT-SAD); and (c) one with the blank ground truth labels (GT-labels). The results show that most human reviews tend to be quite similar, albeit with some outliers, but the choice of GT-labels can make a dramatic difference to scored performance. Using the GT-SAD provides a big advantage and improves human review scores substantially, though small differences in the GT-SAD used can have a dramatic effect on results. The use of forgiveness collars is shown to be unhelpful. The results show that state-of-theart systems can outperform the best human reviews when no prior information is provided. However, the best human reviews still outperform state-of-the-art systems when starting from the GT-SAD.