AIAny
Icon for item

WenetSpeech-Chuan: A Large-Scale Sichuanese Corpus With Rich Annotation For Dialectal Speech Processing

Provides a 10,000-hour Sichuanese (Chuan-Yu) speech corpus with rich annotations (timestamps, speaker age/gender/emotion, SNR, DNSMOS) and unified metadata for ASR and TTS research; includes metadata.jsonl, evaluation benchmarks, and an LLM-assisted transcription pipeline.

Introduction

The lack of large, well-annotated dialectal speech data is a key bottleneck for robust ASR and TTS across regional varieties. This dataset addresses that gap by assembling over 10,000 hours of Sichuanese (Chuan‑Yu) speech from diverse real-world sources and providing multi-dimensional metadata and evaluation splits to enable reproducible research.

What Sets It Apart
  • Scale and coverage: ~10,013 hours spanning short videos, entertainment, live streams, documentaries, audiobooks, drama, interviews, news and more — short videos account for the largest share, boosting real-world diversity.
  • Rich, standardized metadata: single JSONL metadata file with utt_id, timestamps, duration, rover_result, transcription confidence, sample_rate, DNSMOS, SNR/WVMOS, speaker id, age/gender, emotion, domain and original content link — simplifies filtering and segment extraction.
  • Quality tiers and benchmarks: segments labeled as Strong (confidence > 0.90, ~3,714 h) and Weak (0.60–0.90, ~6,299 h) to support supervised and semi-supervised setups; manually verified ASR/TTS eval sets included for fair comparison.
  • Pipeline and transcription methodology: Chuan‑Pipeline automates segmentation, speaker clustering, forced-alignment and multi-modal punctuation; an LLM-GER (LLM Generative Error Correction + ROVER) step merges multiple ASR outputs (including Qwen3 for dialectal normalization) to raise transcription accuracy.
Who it's for and trade-offs

Great fit if you need a large open dialectal speech corpus to train or evaluate ASR/TTS models, study robustness across noisy real-world audio, or develop dialect normalization methods. Look elsewhere if you require fully license-free redistributable audio files (original media links are provided but audio access follows source constraints), extremely clean studio recordings, or very fine-grained speaker demographics beyond the provided age/gender estimates.

Information

  • Websitehuggingface.co
  • OrganizationsAudio, Speech and Language Processing Group (ASLP@NPU), Northwestern Polytechnical University, Beijing AISHELL Technology Co., Ltd., Institute of Artificial Intelligence (TeleAI), China Telecom, School of Intelligence Science and Technology, Nanjing University, WeNet Open Source Community
  • AuthorsYuhang Dai, Ziyu Zhang, Shuai Wang, Longhao Li, Zhao Guo, Tianlun Zuo, Shuiyuan Wang, Hongfei Xue, Chengyou Wang, Qing Wang …
  • Published date2025/09/05

Categories

More Items

Hugging Face

Tracks autonomous research agents publishing studies, discussing ideas, and reviewing one another over twelve weeks. Includes papers, versions, messages, and artifacts; the agents’ reported findings have not been independently verified.

Scales verifiable gameplay data by having a Builder and visual and coding players co-develop web games. Progress-weighted training improves agent performance on both its own benchmark and an external game benchmark.

Hugging Face

Provides executable software-engineering repair tasks for evaluating coding agents. Its 2,955 synthetic instances span 53 Python repositories and include bug-injection patches, reference repairs, and test information.