Detection and handling of overlapping speech for speaker diarization

Abstract

For the last several years, speaker diarization has been attracting substantial research attention as one of the spoken language technologies applied for the improvement, or enrichment, of recording transcriptions. Recordings of meetings, compared to other domains, exhibit an increased complexity due to the spontaneity of speech, reverberation effects, and also due to the presence of overlapping speech. Overlapping speech refers to situations when two or more speakers are speaking simultaneously. In meeting data, a substantial portion of errors of the conventional speaker diarization systems can be ascribed to speaker overlaps, since usually only one speaker label is assigned per segment. Furthermore, simultaneous speech included in training data can eventually lead to corrupt single-speaker models and thus to a worse segmentation. This thesis concerns the detection of overlapping speech segments and its further application for the improvement of speaker diarization performance. We propose the use of three spatial cross-correlationbased parameters for overlap detection on distant microphone channel data. Spatial features from different microphone pairs are fused by means of principal component analysis, linear discriminant analysis, or by a multi-layer perceptron. In addition, we also investigate the possibility of employing longterm prosodic information. The most suitable subset from a set of candidate prosodic features is determined in two steps. Firstly, a ranking according to mRMR criterion is obtained, and then, a standard hill-climbing wrapper approach is applied in order to determine the optimal number of features. The novel spatial as well as prosodic parameters are used in combination with spectral-based features suggested previously in the literature. In experiments conducted on AMI meeting data, we show that the newly proposed features do contribute to the detection of overlapping speech, especially on data originating from a single recording site. In speaker diarization, for segments including detected speaker overlap, a second speaker label is picked, and such segments are also discarded from the model training. The proposed overlap labeling technique is integrated in Viterbi decoding, a part of the diarization algorithm. During the system development it was discovered that it is favorable to do an independent optimization of overlap exclusion and labeling with respect to the overlap detection system. We report improvements over the baseline diarization system on both single- and multi-site AMI data. Preliminary experiments with NIST RT data show DER improvement on the RT ¿09 meeting recordings as well. The addition of beamforming and TDOA feature stream into the baseline diarization system, which was aimed at improving the clustering process, results in a bit higher effectiveness of the overlap labeling algorithm. A more detailed analysis on the overlap exclusion behavior reveals big improvement contrasts between individual meeting recordings as well as between various settings of the overlap detection operation point. However, a high performance variability across different recordings is also typical of the baseline diarization system, without any overlap handling.