2025 GAITE Extras

๐ŸŽ™๏ธ Synthetic Speech Detector

Spot AI-generated speech from spectrograms โ€” the statement literally tells you the solution.

Audio-as-CVResNetGAITEโ˜…โ˜†โ˜†โ˜† friendliest task on this site

๐Ÿ“œThe task, in plain English

Classify speech clips as real or AI-generated. The data comes as Mel spectrograms already prepared as tensors โ€” the audio has been turned into images for you.

InputMel spectrogram tensors
OutputReal vs synthetic (binary)
Really testsWhether you read the statement: it says treat it as image classification, ResNet18 is enough, ~1 epoch

๐Ÿ”งThe baseline you're given

A starter classification pipeline on the spectrogram tensors. The GAITE statement itself supplies the plan โ€” this task is the purest example of "the hints are the intended solution."

๐Ÿš€Baseline vs. solution

โš ๏ธ The baseline (what you are given)
  • Under-trained starter model
  • Spectrogram channels/normalization handled naively
โœ… The winning approach
  • Do what the statement says: ResNet18, adapt input channels, ~1 epoch
  • Normalize spectrograms; check class balance; hold out a validation split
  • Time saved here goes to harder tasks โ€” recognizing an "easy by design" task IS the skill

๐Ÿง’Explain it like I'm brand new

Your Week-3 checkpoint task. It combines everything gently: spectrograms (weird data as images), transfer learning (ResNet18 recipe), a binary metric, and a submission file. When this takes you under two hours end-to-end, your PyTorch literacy is contest-ready.

๐Ÿ’ฌThe Gemma 4 playthrough (2000-token limit)

Chat 1 ยท Full recipe in 3 asks
YOU
Ask 1: dataloaders from tensors X (N,1,128,256), y (N,) โ€” ONLY the Dataset/DataLoader code.
Ask 2 (fresh chat): resnet18 adapted to 1-channel input, 2 classes โ€” ONLY the model surgery.
Ask 3 (fresh chat): training loop, 1 epoch, AdamW, print val accuracy โ€” ONLY the loop.
Three fresh chats, three small pastes, assemble in your notebook. This is the chunked-prompting rhythm to make automatic before contest day.

๐ŸŽฏTakeaways & what Day 1 might do with this

  • Pattern family: spectrograms as images + transfer learning (direct warm-up for 2026 Night Watch).
  • Points-per-hour triage: bank easy tasks fast, spend the surplus where it pays.