This model detects AI-cloned/synthetic voices by analyzing raw audio waveforms through a CNN-Transformer hybrid architecture. It identifies synthetic artifacts that human ears miss: unnatural pitch consistency, GAN-generated frequency smoothness, and missing microtremors.

Architecture

Base: WavLM-base (94.6M params)
Task Head: Sequence Classification (2 classes: bonafide, spoof)
Training: CNN feature extractor frozen, only transformer layers fine-tuned

Training Recipe

Based on WavLM Model Ensemble for Audio Deepfake Detection:

Dataset: ASVspoof 2019 LA (25,380 samples)
Epochs: 5
Effective Batch Size: 64 (16 × 4 gradient accumulation)
Learning Rate: 3e-5 with linear warmup (10%)
Audio Length: 4 seconds (64,000 samples at 16kHz)

Usage

python

from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
import torch, librosa

model_id = "0xmola/wavlm-deepfake-audio-forensics"
extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = AutoModelForAudioClassification.from_pretrained(model_id)
model.eval()

# Load audio
audio, sr = librosa.load("audio.wav", sr=16000)

# Inference
inputs = extractor(audio, sampling_rate=16000, return_tensors="pt", padding=True)
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)
    
# Risk score (0-100, higher = more likely fake)
spoof_idx = model.config.label2id["spoof"]
risk_score = int(probs[0, spoof_idx].item() * 100)
print(f"Risk Score: {risk_score}/100")
print("⚠️ HIGH RISK" if risk_score >= 60 else "✅ LOW RISK")

Demo

Try the live demo: Audio Forensics Deepfake Detector

0xmola/wavlm-deepfake-audio-forensics

WavLM Deepfake Audio Forensics

Model Description

Architecture

Training Recipe

Usage

Demo

References

0xmola/
wavlm-deepfake-audio-forensics