Deep Speaker Embeddings Based Online Diarization

Anastasia Avdeeva, Sergey Novoselov

Abstract: This paper describes our experiments with the Unbounded Interleaved-State Recurrent Neural Network (UIS-RNN) model for development of an online diarization system. For this task several UIS-RNN models based on different speaker embeddings extractors were trained. These systems were evaluated in terms of Diarization Error Rate (DER) metric on public and private test datasets. Also systems were tested on real dialogue data recorded in a bank office. Proposed online models outperform standard offline Agglomerative Hierarchical Clustering (AHC) approach and are compatible with the state-of-the-art Bayesian HMM (VBx) offline method.

External IDs:dblp:conf/specom/AvdeevaN22