Reproduce “A Comparative Study on Multichannel Speaker-Attributed Automatic Speech Recognition in Multi-party Meetings” 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
Mohan Shi, Jie Zhang, Zhihao Du, Fan Yu, Qian Chen, Shiliang Zhang, Li-Rong Dai
This research presents multichannel extensions for three speaker-attributed automatic speech recognition (SA-ASR) architectures, achieving reductions in Speaker-Dependent Character Error Rate (SD-CER) on the AliMeeting corpus by leveraging cross-channel attention mechanisms and neural beamforming to process multi-microphone array data in multi-party meetings.
Key insights
- Leveraging multichannel audio through sophisticated fusion techniques consistently improves SA-ASR performance in multi-party meetings compared to single-channel processing.
- The optimal multichannel fusion strategy is dependent on the specific SA-ASR architecture, with frame-level attention (FLCCA) proving effective for word-level diarization and neural beamforming for joint separation-ASR.
- End-to-end joint training of speech separation and ASR modules can improve overall SA-ASR performance, even if it results in a decrease in intermediate speech separation metrics, suggesting optimization for the ultimate recognition task rather than signal fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.