fassabilf/voiceguard-competition

VoiceGuard — Deepfake Audio Detection Competition

Pelatnas IOAI 2026 | Task 3 of 3

Detect whether a 4-second audio clip is real human speech or AI-generated (TTS/deepfake). Submit probability scores — AUROC is the metric.

Task

Input: .wav audio file (4 seconds, 16 kHz mono)
Output: score — probability (0–1) that the audio is fake
Metric: AUROC (Area Under ROC Curve)

Dataset

Split	Real	Fake	Total
Train	2,874	2,874	5,748
Test	627	627	1,254

• Real: LibriSpeech + VCTK corpus recordings
• Fake: Generated by ≥3 TTS systems (Tacotron2, VITS, SpeechT5); test includes ≥1 unseen TTS system
• Speaker-disjoint: test speakers ≠ train speakers

File Structure

├── train/
│   ├── real/
│   │   └── *.wav
│   └── fake/
│       └── *.wav
├── test/
│   └── *.wav           # flat, unlabeled
├── train.csv           # id, label (real/fake)
├── test.csv            # id (no label)
├── sample_submission.csv   # id, score=0.5
├── solution.csv        # ground truth: id, label
├── notebooks/          # 6 Colab-ready approaches
│   ├── 00_starter.ipynb
│   ├── 01_spectral_lr.ipynb    # AUROC=1.0000
│   ├── 02_acoustic_rf.ipynb    # AUROC=1.0000
│   ├── 03_lfcc_gbdt.ipynb      # AUROC=1.0000
│   ├── 04_cnn_mel.ipynb        # AUROC=1.0000
│   └── 05_rawnet.ipynb         # AUROC=1.0000
├── submissions/
└── writeup/
    └── writeup.md

Submission Format

id,score
test/voiceguard_00001.wav,0.92
test/voiceguard_00002.wav,0.04
...

Score = P(fake). Higher = more likely fake.

How to Load

import pandas as pd, librosa
train_df = pd.read_csv("train.csv")
train_df["label_int"] = (train_df["label"] == "fake").astype(int)
y, sr = librosa.load(f"train/{train_df.iloc[0]['id']}", sr=16000)

Notebooks (Colab-ready)

All approaches achieve AUROC=1.0 — the TTS artifacts are clearly distinguishable. Focus on understanding why each feature works.

Notebook	Approach	Val AUROC
01_spectral_lr	Spectral flatness + HNR + MFCC → LogReg	1.0000
02_acoustic_rf	Full acoustic features → ExtraTrees	1.0000
03_lfcc_gbdt	LFCC + delta → GradientBoosting	1.0000
04_cnn_mel	Mel (80×501) + CNN + BCELoss	1.0000
05_rawnet	Raw waveform + SincConv + GRU	1.0000

Note: AUROC=1.0 means all approaches perfectly separate real from fake on this dataset. In production, use harder datasets like ASVspoof 2021 DF or In-the-Wild.

Why Anti-Spoofing Matters

Voice deepfakes enable voice phishing (vishing), identity fraud, and disinformation. Key benchmarks: ASVspoof 2021, ADD Challenge.

Citation

ASVspoof 2019: A large-scale public database of spoofed and genuine speech.
Wang et al., Computer Speech & Language, 2020.