Reproduce “LibriheavyMix: A 20,000-Hour Dataset for Single-Channel Reverberant Multi-Talker Speech Separation, ASR and Speaker Diarization” on OpenResearch

About this paper

Zengrui Jin, Yifan Yang, Mohan Shi, Wei Kang, Xiaoyu Yang, Zengwei Yao, Fangjun Kuang, Liyong Guo and 5 more

LibriheavyMix is presented as a 20,000-hour single-channel dataset for reverberant multi-talker speech separation, ASR, and speaker diarization, which effectively combines the scale of simulated data with realistic acoustic complexities. Models trained on this dataset achieve a 28.1% cpWER on the WHAMR! 2-speaker test set, representing a substantial improvement over the 61.4% cpWER from WHAMR!-only training.

Key insights

• Large-scale simulated datasets can achieve a high degree of realism by incorporating statistical properties from real-world meeting scenarios and advanced reverberation modeling, enabling better model generalization.
• Increasing the scale of training data for multi-talker tasks, even in simulated reverberant environments, significantly improves performance across speech separation, ASR, and speaker diarization.
• Models trained on LibriheavyMix demonstrate enhanced generalization to external challenging datasets (e.g., WHAMR!), showcasing the dataset's effectiveness in developing broadly applicable speech processing solutions.

Summary by alphaXiv.