Discord

Checklist IOAI 2025 GAITE Contest · Task 5

Synthetic Speech Detector

Train a classifier that tells real human speech from AI-generated speech, using precomputed Mel-spectrogram tensors.

  • Audio
  • Binary classification (synthetic speech detection)

The task

AI-generated speech is now convincing enough to be misused, for example to fake recordings of public figures. The goal is a model that separates synthetic speech from genuine human recordings.

Instead of raw audio, the data comes as Mel spectrograms saved as .pt tensors, which look like 2-D images of time against Mel frequency. Real recordings are labelled bonafide (0) and synthetic ones spoof (1). A provided SpectrogramDataset class loads them and must not be modified.

Predictions are needed for hidden validation and test sets. The statement suggests a ResNet18 or a self-made CNN, notes that the one-epoch baseline is undertrained, and advises not to dwell on how the spectrograms were computed. The submitted notebook may contain just the trained model.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
Mel-spectrogram tensors (.pt) in training_set/bonafide and training_set/spoof; hidden validation and testing sets of unlabelled spectrograms.
You submit
submission.ipynb producing a zip with submissionA.csv (validation) and submissionB.csv (test), one predicted label (0 or 1) per line and no header.
Scoring
F1-score against ground_truth_labels.csv (the repo's metrics.py uses macro-averaged F1). The Scientific Committee's highest score (0.90) and baseline score (0.70) on Leaderboard B were used for score normalisation.
Rules
  • Notebook must finish within 20 minutes on the test machine
  • At most 50 successful submissions; up to 2 can be selected and the higher Leaderboard B score counts
  • Mounted datasets or self-made .pth files may be submitted with the notebook
  • Test machine has no internet access; pretrained models may be used only if they import offline
  • Do not modify the provided SpectrogramDataset loader
Format
GAITE Contest, Day 2 (6 August 2025); six-hour contest day

Details

Year
2025, Beijing, China
Round
GAITE Contest · Task 5
Language
English
License
CC BY 4.0, as stated by the source