How hyeminss11/true-voice-gemma4 works
TrueVoice is a browser-accessible voice deepfake detector, with a separate notebook-based training and evaluation workflow. The documented user flow accepts uploaded or recorded audio, optionally simulates phone-call codec conditions, and returns a real/fake verdict using Gemma’s frozen audio tower and a trained classifier head. The README describes these relationships, but no notebook source was provided, so implementation details beyond that documentation are not verified.
Main entry points: truevoice_demo.ipynb Gradio demo · truevoice_finetune.ipynb Head training · classifier_head.pt Classifier weights · truevoice_data_preprocess_pipeline.ipynb Data preprocessing
Components 18
Detection demo
- Gradio demo
notebooks/truevoice_demo.ipynb - Audio input
notebooks/truevoice_demo.ipynb - Phone simulation
notebooks/truevoice_demo.ipynb - Real / fake verdict
notebooks/truevoice_demo.ipynb
Audio inference
- Gemma audio processor
notebooks/truevoice_demo.ipynb - Frozen audio tower
notebooks/truevoice_demo.ipynb - Mean pooling
notebooks/truevoice_demo.ipynb - Classification head
notebooks/truevoice_demo.ipynb - Classifier weights
saved_model/classifier_head.pt
Training pipeline
- Training metadata
saved_model/metadata.json - Data preprocessing
notebooks/truevoice_data_preprocess_pipeline.ipynb - Audio dataset
notebooks/truevoice_finetune.ipynb - On-the-fly collator
notebooks/truevoice_finetune.ipynb - Head training
notebooks/truevoice_finetune.ipynb
Evaluation
- EER evaluation
notebooks/truevoice_test.ipynb
Other components
- User
- Gemma 4 E4B
- ASVspoof data
Connections 19
- User → Gradio demo: uses
README.md - Gradio demo → Audio input: accepts
README.md - Gradio demo → Phone simulation: offers
README.md - Gradio demo → Gemma audio processor: processes audio
README.md - Gemma audio processor → Frozen audio tower: extracts features
README.md - Frozen audio tower → Mean pooling: provides features
README.md - Mean pooling → Classification head: supplies features
README.md - Classification head → Real / fake verdict: predicts
README.md - Gemma 4 E4B → Gemma audio processor: provides
README.md - Gemma 4 E4B → Frozen audio tower: provides
README.md - Classifier weights → Classification head: loads into
README.md - ASVspoof data → Data preprocessing: supplies audio
README.md - Data preprocessing → Audio dataset: prepares data
README.md - Audio dataset → On-the-fly collator: loads batches
README.md - On-the-fly collator → Head training: provides batches
README.md - Gemma 4 E4B → Head training: provides frozen features
README.md - Head training → Classifier weights: saves head
README.md - Head training → Training metadata: records metadata
README.md - ASVspoof data → EER evaluation: supplies eval set
README.md