PLDA-based mean shift speakers' short segments clustering

Salmun, Itay; Shapiro, Ilya L.; Opher, Irit; Lapidot, Itshak

doi:10.1016/j.csl.2017.04.006

Cited by 13 publications

(10 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Two main options, cosine similarity, which lacks of any knowledge about the speaker subspace, and PLDA likelihood ratio are tested. Efficient clustering techniques are considered, such as AHC and MS [28][29][30][31].…”

Section: Methods For Domain Mismatch Reductionmentioning

confidence: 99%

“…Some approaches make clustering decisions relying on pairwise relationships between representations, such as AHC [10,[24][25][26]. Other approaches make use of relationships among multiple segments in a limited area, e.g., Mean-Shift (MS) [27][28][29][30][31]. Decisions can also be made keeping in mind all the acoustic segments, as Kmeans [7,32], variational Bayes [33], and fully Bayesian PLDAs [8,9,34].…”

Section: Speaker Diarization State Of the Artmentioning

confidence: 99%

See 1 more Smart Citation

Unsupervised adaptation of PLDA models for broadcast diarization

Viñals

Giménez

Villalba

et al. 2019

J AUDIO SPEECH MUSIC PROC.

View full text Add to dashboard Cite

We present a novel model adaptation approach to deal with data variability for speaker diarization in a broadcast environment. Expensive human annotated data can be used to mitigate the domain mismatch by means of supervised model adaptation approaches. By contrast, we propose an unsupervised adaptation method which does not need for in-domain labeled data but only the recording that we are diarizing. We rely on an inner adaptation block which combines Agglomerative Hierarchical Clustering (AHC) and Mean-Shift (MS) clustering techniques with a Fully Bayesian Probabilistic Linear Discriminant Analysis (PLDA) to produce pseudo-speaker labels suitable for model adaptation. We propose multiple adaptation approaches based on this basic block, including unsupervised and semi-supervised. Our proposed solutions, analyzed with the Multi-Genre Broadcast 2015 (MGB) dataset, reported significant improvements (16% relative improvement) with respect to the baseline, also outperforming a supervised adaptation proposal with low resources (9% relative improvement). Furthermore, our proposed unsupervised adaptation is totally compatible with a supervised one. The joint use of both adaptation techniques (supervised and unsupervised) shows a 13% relative improvement with respect to only considering the supervised adaptation.

show abstract

Section: Methods For Domain Mismatch Reductionmentioning

confidence: 99%

Section: Speaker Diarization State Of the Artmentioning

confidence: 99%

Unsupervised adaptation of PLDA models for broadcast diarization

Viñals

Giménez

Villalba

et al. 2019

J AUDIO SPEECH MUSIC PROC.

View full text Add to dashboard Cite

show abstract

“…[15] used non flat weights for the shift calculation with cosine distance. In [12] a PLDA similarity was compared with the cosine distance and showed at the beginning some degradation in performances. However, following the work of [16], when the PLDA was trained over short segments while the TV matrix was still trained using long segments, the PLDA-based mean-shift clustering outperformed the cosine distance-based mean-shift clustering.…”

Section: Algorithm 1 Mean-shift Clustering Algorithmmentioning

confidence: 99%

“…All the methods that do not require knowing the number of speakers in advance, can also be applied to the described clustering problem. However, as we have shown in the past, [2,11,12], meanshift based algorithm with PLDA as a similarity measure performs well for the off-line task, so we decided to continue with this approach and show an easy way to extend it to the described problem of incremental on-line clustering. That way we can explore the limitations of the mean-shift clustering under the new conditions, and compare the on-line incremental clustering performance with the off-line performance.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Incremental On-Line Clustering of Speakers' Short Segments

Aloni-Lavi¹,

Opher²,

Lapidot³

2018

The Speaker and Language Recognition Workshop (Odyssey 2018)

Self Cite

View full text Add to dashboard Cite

This paper deals with clustering of speakers' short segments, in a scenario where additional segments continue to arrive and should be constantly clustered together with previous segments that were already clustered. In realistic applications, it is not possible to cluster all segments every time a new segment arrives. Hence, incremental clustering is applied in an on-line mode. New segments can either belong to existing speakers, therefore, have to be assigned to one of the existing clusters, or they could belong to new speakers and thus new clusters should be formed. In this work we show that if there are enough segments per speaker in the off-line initial clustering process, it constitutes a good starting point for the incremental on-line clustering. In this case, incremental online clustering can be successfully applied based on the previously proposed mean-shift clustering algorithm with PLDA score as a similarity measure and with k-nearest neighbors (kNN) neighborhood selection.

show abstract