| schema_version: 4 |
| system: |
| name: RawTFNet |
| slug: rawtfnet |
| description: 'Lightweight raw-waveform CNN for speech anti-spoofing: sinc-convolution front-end, depthwise-separable |
| Res2Net-SE conv blocks, and a Tf-SepNet (time-frequency separable) classifier (depth=10, width=32). |
| ASVspoof2019 LA pretrained, FP32, deterministic first-64600-sample window (no random crop).' |
| code: https://github.com/swagshaw/RawTFNet-Pytorch |
| checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawTFNet/blob/aa12f0fe2f10cc5278c954175a12c18cc43e3113/Best_RawTFNet_32.pth |
| params_millions: 0.17754 |
| paper: |
| arxiv_id: '2507.08227' |
| url: https://arxiv.org/abs/2507.08227 |
| bibtex: "@article{xiao2025rawtfnet,\n title={RawTFNet: A Lightweight CNN Architecture for Speech\ |
| \ Anti-spoofing},\n author={Xiao, Yang and Dang, Ting and Das, Rohan Kumar},\n journal={arXiv\ |
| \ preprint arXiv:2507.08227},\n year={2025}\n}\n" |
| dataset: |
| id: SpeechAntiSpoofingBenchmarks/PyAra |
| revision: 43f03384ee9ad701a64e0baaa531c8aedd724cd8 |
| split: test |
| scores: |
| eer_percent: 15.761925192090176 |
| n_trials: 201778 |
| n_skipped: 0 |
| artifact: |
| scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawTFNet/resolve/f2e4b0b2320df6106b5c2ddcc4cd0cd746c97e29/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt |
| scores_sha256: 986166ed6d4d2b5a8b649c60003d8abd71d8e6e70b6fad23a947e7be1ab18888 |
| bench_version: speech-spoof-bench==0.3.4 |
| reproduction: |
| reproduced_by: SpeechAntiSpoofingBenchmarks |
| reproduced_at: '2026-06-10' |
| reproduced_bench_version: speech-spoof-bench==0.3.4 |
| match: scoring |
| submitter: |
| hf_username: korallll |
| contact: k.n.borodin@mtuci.ru |
| submitted_at: '2026-06-10' |
| notes: Deterministic first-64600-sample window (no random crop). Full RawTFNet (width=32), Best_RawTFNet_32.pth. |
|
|
