The Gemma 4 Good Hackathon

The Gemma 4 Good Hackathon · Impact Track · Safety & Trust prize

TrueVoice

Built by Shiwon Oh and hyeminss11

A voice-deepfake research prototype with a frozen Gemma audio encoder and a 1.5 MB classification head. Its browser demo requires a notebook session; the claimed 5.2% equal error rate was not independently confirmed.

I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.

The idea worth stealing

The audio tower is a feature extractor. Freeze it and train 1.5 MB.

TrueVoice never fine-tunes Gemma 4. It runs a mel spectrogram through the model’s frozen audio encoder, mean-pools the 1536-dimensional states, and trains a two-layer head to say real or fake. The decision that mattered was in the data: studio-quality training clips were pushed through an 8 kHz downsample-and-upsample round-trip, so the classifier learns the artifacts of synthetic speech rather than the polish of a studio, the one cue a real phone call would strip away.

What we checked

Date checked: .

Public repository at the submission revision, Kaggle writeup, and linked artifacts; code was read, not executed. Deployment checks and limits are described below.

Labels apply to each finding and the evidence described, not to the project as a whole.

Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.

  1. The head is 1.5 MB

    Verified

    classifier_head.pt is 1,578,541 bytes: Linear 1536→256, GELU, Dropout 0.3, Linear 256→2, trained with the audio tower’s requires_grad set to False. The file named metadata.json is a byte-identical copy of it. Source

  2. The codec is a resample

    Verified

    codec_simulate takes 16 kHz audio down to 8 kHz and back with torchaudio. No G.711, AMR, or Opus is involved, which the writeup states accurately. Source

  3. 5.2 percent EER

    Not independently verified

    All four notebooks have their outputs stripped, so the EER, the 96 percent accuracy, and the 148,176-sample test set live only in the README. Reproducing them needs ASVspoof 2021 and an A100. Source

    What we need: Executed evaluation notebooks, the ASVspoof test manifest, and per-sample scores for the reported EER and accuracy.

Under the hood

E4B’s tower, 29-second clips
Features come from google/gemma-4-e4b-it’s audio_tower with mean pooling over time; inputs are clipped at 29 seconds, and a bfloat16 cast in the collator is what fixed the NaNs.
The demo lives 72 hours
Running the last cell of truevoice_demo.ipynb prints a Gradio share link; there is no hosted deployment, so a suspicious call today needs a Colab session first.

Nuance

A sound recipe and a small, checkable artifact, but the repository is four notebooks and a weight file with no tests or requirements, and the training split is 2,580 real clips against 22,800 fake. On-device E2B and non-English coverage are directions the writeup names, not results.

Sources

Last updated: . Editorial updates do not imply a new technical check.

Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.

The official winner announcement credits Shiwon Oh; the writeup uses shiwoni. The second author is credited as hyeminss11; the demo uploader name alone does not confirm a full author name.

Builder? Add context, request a correction, or ask for a re-check →