Deepfake Audio Detection — Wav2Vec 2.0 Fine-tuned
Fine-tuned Wav2Vec 2.0 model for detecting synthetic (deepfake) speech. Trained on ASVspoof 2019 LA. Cross-dataset evaluations on ASVspoof 2021 LA and WaveFake.
Headline Results
| Evaluation | Equal Error Rate (EER) |
|---|---|
| ASVspoof 2019 LA dev (seen attacks A01-A06) | 0.69% |
| ASVspoof 2019 LA eval (unseen attacks A07-A19) | 5.55% |
| ASVspoof 2021 LA eval (codec-degraded) | 9.09% |
| WaveFake (LJSpeech vocoders, mean) | 29.4% |
On ASVspoof 2021 LA the model matches the strongest published baselines (LFCC-LCNN at 9.26%, RawNet2 at 9.50%) without codec-specific training augmentation.
Architecture
- Backbone: facebook/wav2vec2-base (95M params, 12 transformer layers)
- Input: raw waveform at 16 kHz, 4-second windows (64,000 samples)
- Head: mean-pool over time + linear classifier (768 -> 2)
- Stage 1 training: frozen backbone, classifier head only (1,538 trainable params)
- Stage 2 training (this checkpoint): top 2 transformer layers + final LayerNorm unfrozen (~14M trainable params)
Training Details
- Dataset: ASVspoof 2019 LA training partition (25,380 utterances)
- Class weighting: bonafide=4.92, spoof=0.56 (compensates for ~9:1 spoof:bonafide ratio)
- Optimizer: AdamW
- Learning rate: 1e-5 with 10% warmup + linear decay
- Batch size: 16
- Mixed precision: fp16
- Gradient clipping: 1.0
- Epochs: 10 (best at epoch 9)
- Wall clock: 2h 56m on a single T4 GPU
- Best dev EER: 0.69%
Usage
import torch
from huggingface_hub import hf_hub_download
# Download the checkpoint
ckpt_path = hf_hub_download(
repo_id="Sara1708/deepfake-audio-wav2vec2",
filename="stage2_best.pt",
)
# Load using the inference wrapper from the source repo
from src.inference.predict import DeepfakeDetector
detector = DeepfakeDetector(checkpoint_path=ckpt_path, device="cpu")
result = detector.predict("path/to/audio.wav")
print(result)
The full source code, training notebooks, and evaluation scripts are at: github.com/Saracasm/deepfake-audio-detection
Live demo: huggingface.co/spaces/Sara1708/deepfake-audio-detector (Space link will be live after deployment.)
Limitations
- WaveFake performance is poor (~29% EER on LJSpeech-based vocoders). This model was trained only on ASVspoof attack types and does not generalize well to standalone neural vocoder pipelines (HiFi-GAN, MelGAN, WaveGlow, etc.).
- Codec sensitivity: aggressive lossy compression (GSM, PSTN telephone codecs) degrades performance ~6 percentage points relative to uncompressed audio.
- A10 attack family is a known weakness (15.54% EER on this attack alone).
- This is a research artifact, not a production deepfake detector. Real-world deepfakes may use synthesis methods this model has never seen.
Citation
If you use this model, please cite the underlying datasets:
- ASVspoof 2019: Wang et al., 2020. "ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech." Computer Speech & Language.
- ASVspoof 2021: Yamagishi et al., 2021. "ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection."
- WaveFake: Frank & Schonherr, 2021. "WaveFake: A Data Set to Facilitate Audio Deepfake Detection."
- Wav2Vec 2.0: Baevski et al., 2020. "wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations."
Authors
Sara Iqbal (23K-0669) and Areeba Arif (23K-0618). Spring 2026 Deep Learning Project at FAST-NUCES.