Evaluating inference wrappers for Real vs. AI binary classification.
Welcome to your final, and most exciting, challenge in COMP346! It’s time to step up and tackle one of the most pressing issues in the modern music industry: Responsible Music AI: Audio Deepfake Detection. Your core mission? Build a machine learning model capable of distinguishing authentic, human-composed "Real" music from synthetic "AI" generations.
Armed with a representative subset of the SONICS dataset (straight from MIREX 2025), you'll put your entire music intelligence toolkit to the test. You'll drive the whole pipeline yourself—from digging into Exploratory Data Analysis and extracting key audio features (like MFCCs, Chroma, and Spectrograms) to training and fine-tuning competitive ML models. This leaderboard is the ultimate proving ground to see exactly how your classifier stacks up against the rest of the cohort. Good luck!
While CA2 covers the entire pipeline, this leaderboard specifically evaluates Part 5: Challenge Submission.
To take part in the challenge and get featured on this portfolio-enhancing leaderboard, you must implement the predict_audio(path) wrapper exactly according to the specification. Failure to provide this functioning wrapper means your best model cannot be evaluated blindly against the hidden test set.
Links and materials required to complete the assignment and challenge submission.
Ranked by Final F1-Score, then Accuracy following any penalty. Blind evaluation performed against our hidden test set.
| Rank | Student | Accuracy | Final F1-Score |
|---|---|---|---|
| #1 | Jennifer Pastrana-Dix | 85% | 0.8235 |
| #2 | Albandari Ayob Y M Arab | 80% | 0.7500 |
| #2 | Radek Zienkiewicz | 80% | 0.7500 |
| #3 | Morgan Beynon | 70% | 0.7500 |
| #4 | Kyla Boyce | 80% | 0.7125 (-5% Penalty) |
| #5 | Joe Carruthers | 60% | 0.6923 |