Real-World AVSE Challenge (ISCSLP 2026) — Baseline (track1)

AV-ConvTasNet baseline checkpoint for the Real-World AVSE Challenge, saved with PyTorchModelHubMixin (config.json + model.safetensors). The trained video encoder is bundled in the weights, so no separate lip-reading backbone is needed.

Usage

from look2hear.models import AV_ConvTasNet
model = AV_ConvTasNet.from_pretrained("JusperLee/Real-World-AVSE-Baseline-Track1").eval()

Or directly in the challenge evaluation:

python eval_real.py --ckpt JusperLee/Real-World-AVSE-Baseline-Track1 \
  --track track1 --split dev --metrics all --mode both --save_dir enhanced_out

This repo is private: log in (huggingface-cli login or HF_TOKEN) and be granted access before downloading.

See the baseline repository for the model definition and full pipeline.