SMIIP-lab/AISHELL8-RealScene

📘 Dataset Summary

Property	Description
Language	Chinese (Mandarin, ZH)
License	CC BY-NC-SA 4.0
Total Duration	102.19 hours
Speakers	171 foreground speakers
Scenes	5 real-world locations
Recording Style	Conversational speech
Audio	Near-field + far-field
Video	Multi-view RGB facial video
Sampling Rate	16 kHz
Far-field Audio	8-channel
Video Resolution	256×256 @ 25 fps

---

🎙️ Dataset Description

AISHELL8-RealScene is a public multi-scenario and multi-view audio-visual conversational corpus recorded in real-world environments.

AISHELL8-RealScene focuses on realistic conditions including:

• environmental noise
• viewpoint variation
• speaker interference
• background activities
• outdoor scenes
• conversational interactions

The corpus was collected by AISHELL using a synchronized multi-device recording setup.

---

🏛️ Recording Environment

AISHELL8-RealScene contains recordings from five real-world locations:

ID	Location	Type
L1	Building	Outdoor
L2	Hall	Indoor
L3	Hotel	Indoor
L4	Park	Outdoor
L5	Street	Outdoor

The recordings cover both indoor and outdoor scenarios.

---

🔊 Audio Specifications

AISHELL8-RealScene provides both:

• near-field audio
• far-field microphone-array audio

Near-field Audio

Each foreground speaker wears a close-talking microphone used for transcription and alignment.

Far-field Audio

Far-field audio is recorded using:

• circular microphone array
• 16 microphones
• 16 kHz
• 16-bit

To reduce release size while preserving spatial information, the public release provides:

• 8-channel far-field audio

obtained by uniformly selecting microphones from the array.

Audio Format

• Format: .wav
• Sampling rate: 16 kHz
• Near-field: single-channel
• Far-field: 8-channel

---

🎥 Video Specifications

Video is captured using three synchronized HD RGB cameras.

Camera Views

The three cameras are positioned horizontally with approximately 30° angular intervals:

• D0 — left
• D1 — center
• D2 — right

Video Parameters

Property	Value
Resolution	256×256
Frame Rate	25 fps
Views	D0 / D1 / D2

The released facial videos are:

• 256×256 resolution
• synchronized with audio

---

📊 Dataset Statistics

Dataset Split

Split	Duration (h)	Sessions	Groups	Speakers
Train	79.84	133	56	133
Dev	10.70	18	7	18
Eval	11.65	20	7	20
Total	102.19	171	70	171

No speakers overlap across:

• train
• dev
• eval

All splits contain recordings from all five locations.

---

📜 License

This dataset is released under:

CC BY-NC-SA 4.0

https://creativecommons.org/licenses/by-nc-sa/4.0/

---

📄 Dataset Resources

• 📄 Paper: arXiv 2606.05763

---

📜 Citation

If you use AISHELL8-RealScene, please cite:

@misc{su2026m2savsrmodalityawaremultiviewselfsupervised,
      title={M2S-AVSR: Modality-aware Multi-view Self-supervised Representation for Robust Audio-Visual Speech Recognition}, 
      author={Fei Su and Cancan Li and Ming Li and Juan Liu},
      year={2026},
      eprint={2606.05763},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2606.05763}, 
}