| tags: | |
| - audio | |
| - speech-deepfake-detection | |
| - pytorch | |
| - fairseq | |
| base_model: facebook/wav2vec2-xls-r-300m | |
| # LaST-Net | |
| Model weights for **LaST-Net: Length-Aware Layer and Scale-Adaptive Temporal Network for Speech Deepfake Detection**. | |
| **Training and inference code:** [JunXue-tech/LaST-Net](https://github.com/JunXue-tech/LaST-Net). | |
| ## Checkpoint | |
| `best.pt` is the original epoch-52 checkpoint selected by the lowest mean development EER across 1, 2, 4 and 6 seconds on ASVspoof 2019 LA. It includes the fine-tuned XLS-R 300M frontend, LaST-Net backend, optimizer state and original training metadata. | |
| ## Usage | |
| Follow the environment setup in the [code repository](https://github.com/JunXue-tech/LaST-Net). From that repository: | |
| ```bash | |
| python download_model.py | |
| python infer.py example.wav --checkpoint checkpoints/best.pt \ | |
| --ssl-path /path/to/xlsr2_300m.pt --seconds 6 | |
| ``` | |
| The model constructor requires the original fairseq-format XLS-R 300M checkpoint, available from the [official XLS-R repository](https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec/xlsr), before loading the fine-tuned parameters from `best.pt`. | |
| Input audio must be mono at 16 kHz. The supplied inference code evaluates 1–6 second inputs using prefix cropping and repetition of shorter recordings. Higher `bonafide_log_score` values favor bona fide speech. Scores are not calibrated probabilities. | |
| ## Evaluation | |
| Duration-averaged EER (%) across 1–6 second inputs: 19LA **1.29**, 21LA **4.98**, 21DF **3.62**, and In-the-Wild **7.58**. Per-duration results and evaluation commands are provided in the code repository. Results depend on the evaluation protocol and preprocessing. | |
| ## Author | |
| Jun Xue — [junxue@whu.edu.cn](mailto:junxue@whu.edu.cn) | |