Cross-Channel Attention-Based Target Speaker Voice Activity Detection: Experimental Results for M2MeT Challenge

TL;DR

This paper introduces a cross-channel attention-based target speaker voice activity detection system for multi-channel meeting transcription, significantly reducing diarization error rate and achieving top performance in the M2MeT challenge.

Contribution

The paper proposes a novel cross-channel self-attention mechanism for target speaker VAD, improving multi-channel diarization accuracy over previous methods.

Findings

Single-channel TS-VAD reduces DER by over 75%.

Multi-channel TS-VAD further reduces DER by 28%.

Achieved 1st place in the M2MeT challenge with a DER of 2.98%.

Abstract

In this paper, we present the speaker diarization system for the Multi-channel Multi-party Meeting Transcription Challenge (M2MeT) from team DKU_DukeECE. As the highly overlapped speech exists in the dataset, we employ an x-vector-based target-speaker voice activity detection (TS-VAD) to find the overlap between speakers. For the single-channel scenario, we separately train a model for each of the 8 channels and fuse the results. We also employ the cross-channel self-attention to further improve the performance, where the non-linear spatial correlations between different channels are learned and fused. Experimental results on the evaluation set show that the single-channel TS-VAD reduces the DER by over 75% from 12.68\% to 3.14%. The multi-channel TS-VAD further reduces the DER by 28% and achieves a DER of 2.26%. Our final submitted system achieves a DER of 2.98% on the AliMeeting test set, which ranks 1st in the M2MET challenge.