Reproduce “M2MeT: The ICASSP 2022 Multi-Channel Multi-Party Meeting Transcription Challenge” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fan Yu, Shiliang Zhang, Yihui Fu, Lei Xie, Siqi Zheng, Zhihao Du, Weilong Huang, Pengcheng Guo and 4 more
The paper introduces the AliMeeting corpus, a 120-hour real-recorded Mandarin meeting dataset, and establishes the ICASSP 2022 M2MeT Challenge to advance multi-channel multi-party meeting transcription research. Baseline results demonstrate the dataset's complexity, with an end-to-end multi-speaker ASR model achieving a 29.7% Character Error Rate on beamformed far-field data.
Key insights
- The AliMeeting dataset presents significant challenges, particularly high speaker overlap (average 42.27%) and far-field conditions, leading to higher Diarization Error Rates (DER) in more complex scenarios.
- Data augmentation techniques, specifically simulating noise and reverberation for near-field training data, are crucial for improving far-field ASR performance, reducing CER by 1.7% for single-speaker ASR baselines.
- While beamforming did not improve baseline speaker diarization, it proved beneficial for multi-speaker ASR by enhancing the primary speaker's signal, reducing CER to 29.7% when combined with Serialized Output Training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.