Model card: Arena badges for all 24 datasets
Browse files
README.md
ADDED
|
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- audio
|
| 5 |
+
- anti-spoofing
|
| 6 |
+
- audio-deepfake-detection
|
| 7 |
+
- speech
|
| 8 |
+
- asvspoof
|
| 9 |
+
- multi-view
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# AMSDF
|
| 13 |
+
|
| 14 |
+
<!-- ssb:badges -->
|
| 15 |
+
|
| 16 |
+
**Audio Multi-view Spoofing Detection Framework (AMSDF)** (Wu et al., IEEE TIFS
|
| 17 |
+
2024). AMSDF captures intra- and inter-view spoofing cues from **audio–text–emotion
|
| 18 |
+
correlations**: an XLS-R (wav2vec2-xls-r-300m) **text view**, an ACRNN **emotion
|
| 19 |
+
view** (logfbank + delta + delta²), and a raw-audio **graph view**, fused by
|
| 20 |
+
heterogeneous graph attention with a group-based readout. The model takes a raw
|
| 21 |
+
speech waveform (and a derived emotion spectrogram) and returns a score where
|
| 22 |
+
**higher = more bona fide**.
|
| 23 |
+
|
| 24 |
+
- **Original code:** https://github.com/ItzJuny/AMSDF
|
| 25 |
+
- **Paper:** https://ieeexplore.ieee.org/document/10605761 (IEEE TIFS 2024)
|
| 26 |
+
- **Original authors:** Junyan Wu, Qilin Yin, Ziqi Sheng, Wei Lu, Jiwu Huang, Bin Li
|
| 27 |
+
- **Parameters:** ~325.4 M
|
| 28 |
+
- **Checkpoint:** [`M1.pth`](./M1.pth) (the authors' best-EER seed; M2/M3 also released)
|
| 29 |
+
|
| 30 |
+
This repo is self-contained for inference: the multi-view network is vendored in
|
| 31 |
+
[`models/`](./models), the builder in [`_net.py`](./_net.py), and the exact wrapper
|
| 32 |
+
in [`amsdf.py`](./amsdf.py).
|
| 33 |
+
|
| 34 |
+
## Arena results
|
| 35 |
+
|
| 36 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 37 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 38 |
+
|
| 39 |
+
EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under `.eval_results/`.
|
| 40 |
+
|
| 41 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 42 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 43 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 44 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 45 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 46 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 47 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 48 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 49 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 50 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 51 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 52 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 53 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 54 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 55 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 56 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 57 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 58 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 59 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 60 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 61 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 62 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 63 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 64 |
+
[](https://huggingface.co/spaces/SpeechAntiSpoofingBenchmarks/SpeechAntiSpoofingArena?system=amsdf)
|
| 65 |
+
|
| 66 |
+
## Architecture & inference notes
|
| 67 |
+
|
| 68 |
+
- **Two inputs** computed by the wrapper from each waveform: the raw waveform
|
| 69 |
+
(text + audio views) and an emotion spectrogram `(3, 300, 40)` =
|
| 70 |
+
logfbank(40) + Δ + Δ² (first 300 frames), as in the upstream `data_io.extract_mel`.
|
| 71 |
+
- **Score = `P(real) − P(fake)`** (`out[:,0] − out[:,1]`). The repo's label
|
| 72 |
+
convention is class 0 = real, class 1 = fake, so the bonafide score is the
|
| 73 |
+
*negative* of the repo's reported `out[:,1]`.
|
| 74 |
+
- Input: raw float32 mono @ 16 kHz, first-**64600**-sample (≈4 s) window
|
| 75 |
+
(tile-repeat if shorter). FP32.
|
| 76 |
+
- Requires the public XLS-R `xlsr2_300m.pt` and the emotion model `ser_acrnn.pth`
|
| 77 |
+
in `pretrained_models/` (loaded at construction; the M*.pth checkpoint then
|
| 78 |
+
overwrites the fine-tuned weights). The fairseq XLS-R is loaded via
|
| 79 |
+
`merge_with_parent` (cross-fairseq-version fix).
|
| 80 |
+
|
| 81 |
+
## Credits
|
| 82 |
+
|
| 83 |
+
All model design and weights are the work of the original AMSDF authors
|
| 84 |
+
(https://github.com/ItzJuny/AMSDF). This repository only packages their released
|
| 85 |
+
checkpoint for the Speech Anti-Spoofing Arena.
|
| 86 |
+
|
| 87 |
+
```bibtex
|
| 88 |
+
@article{wu2024audio,
|
| 89 |
+
author = {Wu, Junyan and Yin, Qilin and Sheng, Ziqi and Lu, Wei and Huang, Jiwu and Li, Bin},
|
| 90 |
+
journal = {IEEE Transactions on Information Forensics and Security},
|
| 91 |
+
title = {Audio Multi-view Spoofing Detection Framework Based on Audio-Text-Emotion Correlations},
|
| 92 |
+
year = {2024}, volume = {19}, pages = {7133--7146},
|
| 93 |
+
doi = {10.1109/TIFS.2024.3431888}
|
| 94 |
+
}
|
| 95 |
+
```
|