Reproduce “Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context” 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

Wei Kang, Xiaoyu Yang, Zengwei Yao, Fangjun Kuang, Yifan Yang, Liyong Guo, Long Lin, Daniel Povey

Researchers at Xiaomi Corp. released Libriheavy, a 50,000-hour English ASR corpus featuring complete punctuation, casing, and textual context. This dataset facilitates the development of advanced end-to-end ASR systems capable of producing full-format text and leveraging cross-utterance information.

Key insights

  • The performance gap between ASR models trained on normalized text and those trained on richly formatted text (with punctuation and casing) significantly diminishes and becomes negligible with sufficiently large training datasets (e.g., 50,000 hours).
  • A general, open-source audio alignment pipeline can efficiently and accurately process tens of thousands of hours of unlabeled audio to create high-quality, richly annotated ASR corpora.
  • Consistent improvements in ASR performance with increasing data volume (from 500 to 50,000 hours) confirm the critical role of large-scale supervised data for robust end-to-end neural ASR models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.