Abstract

DIHARD challenge focuses on the hard diarization problem and the DIHARD dataset includes a number of challenging domains that are hard to obtain low diarization error rates. We propose a novel approach to deal with domain mismatch problems by estimating the domain of the given input session. We take advantage of three different embedding extractors trained on different datasets. Based on these multiple embedding extractors, our domain adaptive speaker diarization system employs two different approaches: Hard decision and soft decision. In the hard decision method, we estimate the given session into one of the three categories and select an embedding extractor suited to that category. On the other hand, in the soft decision method, we train our proposed neural affinity score fusion network that estimates the desirable weights for the affinity scores we obtain from the three embedding extractors. We show the performance gain from each method and how our domain estimator models are trained to obtain such improvement. In addition, we introduce the auto-tuning spectral clustering method to develop a parameter-free diarization system.

Date

2021

Authors

Tae Jin Park, Raghuveer Peri, Arindam Jati, Shrikanth Narayanan

Journal

Proc. 3rd DIHARD Speech Diarization Challenge Workshop