Discord

Checklist NEOAI 2026 Day 2 · Task 4

Day 2 Audio (Sped-up Speech Detection)

Detect which test speech clips have been sped up, using only a supplied Whisper checkpoint.

  • Audio
  • Anomaly detection / binary scoring
Read the official statement

The task

The data are 7-second, 16 kHz mono WAV recordings from multiple speakers. The training split contains only normal-speed clips (about 30 per speaker); the test split (about 10 clips per speaker) mixes normal and sped-up clips. Both splits share the same speakers and the speaker_id is given.

A self-contained Whisper inference library (whisper_lib) is shipped with the data. It performs greedy decoding and returns the transcript, the clip duration and per-step token log-probabilities, and also exposes the encoder. The baseline scores each clip by how much lower its average token log-probability is than the training mean.

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
train.csv and test.csv (audio_id, speaker_id, audio_path), WAV files, whisper_lib and the Whisper checkpoint in whisper_assets/.
You submit
CSV with columns audio_id, score (higher means more likely sped up).
Scoring
ROC-AUC.
Rules
  • The only pretrained model allowed is the Whisper checkpoint provided in the environment (whisper_assets/).
  • General NEOAI 2026 rules apply.
Format
NEOAI 2026 main round, Day 2 (4 May 2026). Individual.

Notebooks

Competition

Source files

Details

Year
2026, SberUniversity, Russia; online and national venues for other countries
Round
Day 2 · Task 4
Language
English
License
Per-competition Kaggle data licence: MIT (Task 4); Subject to Competition Rules (Tasks 5 and 6), as stated by the source