saracen9/lr-asd-talkset

LR-ASD finetuning_TalkSet.model (unchanged copy)

Status: public copy, uploaded on 2026-11-12 to saracen9/lr-asd-talkset by the owner of curu-vision (curu-vision spec 023, research R1.5: the owner's choice of the weights source).

This repository holds an unchanged copy of the file weight/finetuning_TalkSet.model of the LR-ASD repository. The copy has the same bytes as the file at the original download URL. curu-vision loads it in the active-speaker port, through its pin lr-asd-talkset.

The file

Field	Value
File	finetuning_TalkSet.model (weight/finetuning_TalkSet.model in the original repository)
Size	3,426,337 bytes
SHA-256	6b4ef53694e874e96cf630198dc479c78aebb3993bbf166aee3d926dfe7d9342
Downloaded	2026-11-12
Change	None. The file is an unchanged copy of the original file.

The original source

Field	Value
Original repository	Junhua-Liao/LR-ASD
Revision	1b6dcd2d8fc2895683de6508ec6294ec47d388ca (2025-03-23)
Original download URL	https://raw.githubusercontent.com/Junhua-Liao/LR-ASD/1b6dcd2d8fc2895683de6508ec6294ec47d388ca/weight/finetuning_TalkSet.model
Paper	J. Liao, H. Duan, K. Feng, W. Zhao, Y. Yang, L. Chen and Y. Chen, "LR-ASD: Lightweight and Robust Network for Active Speaker Detection", International Journal of Computer Vision, 2025 (PDF)
Earlier paper	J. Liao, H. Duan, K. Feng, W. Zhao, Y. Yang and L. Chen, "A Light Weight Model for Active Speaker Detection", CVPR, 2023, pages 22932 to 22941 (PDF)
Authors	Junhua Liao, Haihan Duan, Kanghui Feng, Wanbing Zhao, Yanbing Yang, Liangyin Chen and Yanru Chen

Licence and terms

• Code licence: MIT, "Copyright (c) 2025 Liao Junhua" (LICENSE).
• Terms of the weights: the repository states no separate terms for the files in weight/. The file is in the repository that the MIT licence covers. The README says: "Please cite our papers if you use this code or model weights." (README.md).
• The owner of curu-vision accepted the TalkSet weights on 2026-11-12 (curu-vision spec 023, research R1.4: the owner's choice of the weights).

Training data

• The released model: the README says: "We have also provided the model weights fine-tuned on the TalkSet dataset." The README does not say which model was fine-tuned.

• TalkSet: the README of LR-ASD does not describe TalkSet. The TalkNet paper (Tao et al., 2021, section 4.1.3) builds TalkSet from clips of VoxCeleb2 and LRS3, and does not publish it.

• The other weights of the repository: the README says that weight/pretrain_AVA.model holds "The model weights trained on the AVA dataset". That file is not this file.

• This copy: it does not change the terms of the original file or of its training data. Check those terms, and the terms of the training datasets, before you use the file.

Use in curu-vision

• The active-speaker port of curu-vision (load_speaker_detector) loads the file into its own copy of the network, with the MIT notice. The copy follows model/Model.py, model/Encoder.py, model/Classifier.py and loss.py.
• The integration tier of curu-vision gives each frame score within 1.2e-7 of the LR-ASD code on the same crops of a demo clip.

Citation

@article{Liao_2025_IJCV,
  title     = {LR-ASD: Lightweight and Robust Network for Active Speaker Detection},
  author    = {Liao, Junhua and Duan, Haihan and Feng, Kanghui and Zhao, Wanbing and Yang, Yanbing and Chen, Liangyin and Chen, Yanru},
  journal   = {International Journal of Computer Vision},
  pages     = {1--21},
  year      = {2025},
  publisher = {Springer}
}