Rezzsl/Real-World_AVSE_Challenge_dev_and_test_sets_ISCSLP_2026

Real-World AVSE Challenge (ISCSLP 2026)
Development & Test Sets

The Real-World Audio-Visual Speech Enhancement (AVSE) Challenge is an ISCSLP 2026 Special Session that pushes audio-visual speech enhancement beyond simulated-speech, clean-video benchmarks, toward natural speech overlap, real rooms and unreliable video.

This repository releases the development set and the test set with ground truth used in the challenge. Official test-set results are on the leaderboard.

Track	Theme
Track 1	Real-World Mixed Scenarios: naturally captured multi-speaker audio-visual data.
Track 2	Visual Degradation: target-speaker video under degraded visual conditions (full mask, occlusion, low resolution, frame freeze, audio-visual desynchronization).

Files

File	Size	Contents
track1_dev.zip	1.5 GB	Track 1 development set
track2_dev.zip	1.7 GB	Track 2 development set
Testing_set_GT.zip	6.1 GB	Track 1 & Track 2 test sets, with ground truth

Statistics

Track	Split	mix clips	remix clips	Hours (mix + remix)
Track 1	dev	1,242	900	1.9 + 1.6
Track 1	test	2,472	1,785	3.7 + 2.9
Track 2	dev	1,527	1,098	2.4 + 2.0
Track 2	test	2,820	2,121	4.3 + 3.5

The data covers seven groups of speakers, two speakers per group. Three groups form the development set and the remaining four form the test set, so dev and test share no speakers. Every clip contains two speakers (s1, s2), and each one is a separate target.

Scenes

Each split contains two scenes:

• mix: real recordings of two people talking at the same time. mix.wav is a real mixture, so there is no clean reference.
• remix: two clean single-speaker clips mixed at an SNR between −5 and 5 dB. The clean sources ship as s1.wav / s2.wav.

Directory structure

Development set (track1_dev.zip, track2_dev.zip):

dev/
├── mix/<mix_id>/               mix.wav  s1.mp4 s2.mp4  s1.pkl s2.pkl  s1.txt s2.txt
├── remix/<s1_id>_<s2_id>/      mix.wav  s1.mp4 s2.mp4  s1.pkl s2.pkl  s1.txt s2.txt  s1.wav s2.wav
├── mix_manifest.json
└── remix_manifest.json

Test set (Testing_set_GT.zip):

Testing_set_GT/Real-avse/
├── track1/test/
│   ├── mix/<mix_id>/               mix.wav  s1.mp4 s2.mp4  s1.pkl s2.pkl  s1.txt s2.txt
│   ├── remix/<s1_id>_<s2_id>/      mix.wav  s1.mp4 s2.mp4  s1.pkl s2.pkl  s1.txt s2.txt  s1.wav s2.wav
│   ├── mix_manifest.json
│   ├── remix_manifest.json
│   └── blind_map.json
└── track2/test/
    ├── mix/  remix/                (same layout as track1)
    ├── mix_manifest.json
    ├── remix_manifest.json
    ├── mix_perturb_manifest.json
    ├── remix_perturb_manifest.json
    └── blind_map.json

Per-clip files

File	Description
mix.wav	Two-speaker mixture. 16 kHz, mono, 16-bit PCM.
s1.mp4, s2.mp4	Face video of each speaker. 256×256, 25 fps, H.264, time-aligned with mix.wav.
s1.pkl, s2.pkl	Precomputed 68-point face landmarks: a Python list with one entry per video frame, each a (68, 2) float32 array, or None if no face was detected. Saved with NumPy 2.x, so loading requires numpy>=2. Not provided for some Track 2 clips.
s1.txt, s2.txt	Mandarin transcript of each speaker (UTF-8).
s1.wav, s2.wav	Clean reference of each speaker (remix only). 16 kHz, mono.

Manifests

mix_manifest.json / remix_manifest.json contain one entry per clip:

Field	Description
mix_id	Clip ID (the folder name). For remix, <s1_id>_<s2_id>.
s1_speaker, s2_speaker	Speaker labels, <group>_<L/R>.
s1_id, s2_id	(remix) IDs of the source clips.
split, dir	(test) Split name and clip path relative to the track folder.
group, s1_side, s2_side	(test, mix) Recording group and speaker side (L / R).
snr_db	(test, remix) Mixing SNR in dB.
s1_start_frame, s2_start_frame, n_frames, length_samples	(test, remix) Frame offsets of s1 / s2, length in video frames and in audio samples.

mix_perturb_manifest.json / remix_perturb_manifest.json (Track 2 test) list the visual degradation applied to each face:

Field	Description
mix_id, kind, dir	Clip ID, scene (mix / remix), clip path.
faces.s1, faces.s2	Degradation for each speaker's video.
deg_type / deg_name	0 FULL_MASK, 1 OCCLUSION, 2 LOW_RESOLUTION, 3 FRAME_FREEZE, 4 AV_DESYNC.
mask_start_frame, mask_len_frames	Start frame and length of the degraded segment.
shift_frames	Audio-visual offset in frames (AV_DESYNC).
n_frames	Number of video frames in the clip.

blind_map.json maps the anonymized IDs (blind_id) used in the challenge-period test release to the clip IDs here (scene, mix_id), so you can score submissions made during the challenge against this ground truth.

Download

hf download Rezzsl/Real-World_AVSE_Challenge_dev_and_test_sets_ISCSLP_2026 --repo-type dataset --local-dir Real-World-AVSE
cd Real-World-AVSE
unzip -q track1_dev.zip -d track1      # -> track1/dev/
unzip -q track2_dev.zip -d track2      # -> track2/dev/
unzip -q Testing_set_GT.zip            # -> Testing_set_GT/Real-avse/track{1,2}/test/

Both dev archives have dev/ as their top-level folder. Extract them into separate folders as shown above, or the second one will overwrite the first.

Usage

import pickle
import soundfile as sf

d = "Testing_set_GT/Real-avse/track1/test/remix/000129_000662"
mix, sr = sf.read(f"{d}/mix.wav")                     # 16 kHz mixture
ref, _ = sf.read(f"{d}/s1.wav")                       # clean reference of speaker 1
text = open(f"{d}/s1.txt", encoding="utf-8").read().strip()
landmarks = pickle.load(open(f"{d}/s1.pkl", "rb"))    # list[(68, 2) array | None], one per frame

The baseline repository provides a ready-made data loader (look2hear/datas/real_test_dataset.py) and the full evaluation pipeline (eval_real.py).

Evaluation

Metric	Type	Scenes
SI-SDR / PESQ / STOI	reference-based	remix
UTMOS	no-reference MOS	mix + remix
DNSMOS (P.808 / SIG / BAK / OVRL)	no-reference	mix + remix
CER (Fun-ASR-Nano)	intelligibility	mix + remix
Speaker similarity (WeSpeaker ResNet34)	identity preservation	mix + remix

The official ranking uses OVRL, the mean of a team's per-metric ranks (lower is better). See the leaderboard for test-set results and the baseline repository for the evaluation code and baseline scores.

Organizers

Kai Li (Tsinghua University) · Wenze Ren (National Taiwan University) · Junjie Li (The Hong Kong Polytechnic University) · Cheng Yu (Ohio State University) · Peijun Yang (Wuhan University) · Chien-yu Huang (Carnegie Mellon University) · Haibin Wu (Meta) · Szu-Wei Fu (NVIDIA) · Wen-Chin Huang (Nagoya University) · Hsin-Min Wang (Academia Sinica) · Xiaolin Hu (Tsinghua University) · Ming Li (CUHK-Shenzhen) · DeLiang Wang (CUHK-Shenzhen) · Yu Tsao (Academia Sinica)

Website: https://real-world-avse.github.io/ · Contact: realworldavse_iscslp2026@googlegroups.com

Citation

If you use this dataset, please cite:

@article{li2026iscslp,
  title={The ISCSLP 2026 Real-World Audio-Visual Speech Enhancement Challenge},
  author={Li, Kai and Ren, Wenze and Li, Junjie and Yu, Cheng and Yang, Peijun and Huang, Chien-yu and Wu, Haibin and Fu, Szu-Wei and Huang, Wen-Chin and Wang, Hsin-Min and others},
  journal={arXiv e-prints},
  pages={arXiv--2608},
  year={2026}
}