Chinese-LiPS: A Chinese audio-visual speech recognition dataset with Lip-reading and Presentation Slides

⭐ Introduction

The Chinese-LiPS dataset is a multimodal dataset designed for audio-visual speech recognition (AVSR) in Mandarin Chinese. This dataset combines speech, video, and textual transcriptions to enhance automatic speech recognition (ASR) performance, especially in educational and instructional scenarios.

🚀 Dataset Details

• Total Duration: 100.84 hours
• Number of Speakers: 207 professional speakers
• Number of Clips: 36,208 video clips
• Audio Format: Stereo WAV, 48 kHz sampling rate
• Video Format:
  • Slide Video: 1080p resolution, 30 fps
  • Lip-Reading Video: 720p resolution, 30 fps
• Annotations: JSON format with transcriptions and extracted text from slides

Dataset Statistics

Split	Duration (hrs)	# Segments	# Speakers
Train	85.37	30,341	175
Validation	5.35	1,959	11
Test	10.12	3,908	21
Total	100.84	36,208	207

📂 Dataset Organization

The dataset is structured into several compressed files:

• image.zip: First-frame images from slide videos (used for OCR and vision-language models).

• processed_test.zip processed_val.zip processed_train.zip: Processed data with 16 kHz audio, 96×96 25-frame lip-reading videos, and JSON annotations.

• train.zip, test.zip, val.zip: Data split into training, testing, and validation sets. Each contains:

  ├── ID1_age_gender_topic/
  │   ├── WAV/
  │   │   ├── ID1_age_gender_topic_001.json  # Annotation file
  │   │   ├── ID1_age_gender_topic_001.wav   # Audio file (48 kHz)
  │   ├── PPT/
  │   │   ├── ID1_age_gender_topic_001_PPT.mp4  # Slide video (1080p 30fps)
  │   ├── FACE/
  │   │   ├── ID1_age_gender_topic_001_FACE.mp4  # Lip-reading video (720p 30fps)
  ├── ...
  

• meta_all.csv, meta_train.csv, meta_valid.csv, meta_test.csv: Metadata files with ID, TOPIC, WAV, PPT, FACE, and TEXT fields.

  The TOPIC field is abbreviated in Chinese as follows: DZJJ = E-sports & Gaming, JKYS = Health & Wellness, KJ = Science & Technology, LY = Travel & Exploration, QC = Automobile & Industry, RWLS = Culture & History, TY = Sports & Competitions, YS = Movies & TV Series, ZX = Others.

• meta_test.json: Includes OCR and InternVL2 prompts for the test set.

  wav_path: Path to the audio file.
  ppt_path: Path to the first-frame image of the slide video.
  ocr_text: Text extracted by PaddleOCR.
  vl2_text: Text extracted by InternVL2.
  gt_text: Ground truth transcription of the audio.
  ocr_vl2_text: OCR text reprocessed by InternVL2 (not a concatenation of PaddleOCR and InternVL2 results).
  

📥 Download

You can download the dataset from the following sources:

• Download from OneDrive
• Download from Huggingface
• Download from Baidu Netdisk (Password: vg2a)

📚 Citation

@misc{zhao2025chineselipschineseaudiovisualspeech,
  title={Chinese-LiPS: A Chinese audio-visual speech recognition dataset with Lip-reading and Presentation Slides}, 
  author={Jinghua Zhao and Yuhang Jia and Shiyao Wang and Jiaming Zhou and Hui Wang and Yong Qin},
  year={2025},
  eprint={2504.15066},
  archivePrefix={arXiv},
  primaryClass={cs.MM},
  url={https://arxiv.org/abs/2504.15066}
}