Datasets:
Dataset Description
Overview
The dataset contains 7,002 four-second WAV recordings of English speech — half from real human speakers, half generated by neural TTS systems.
- Train: 5,748 clips (2,874 real, 2,874 fake)
- Test: 1,254 clips (627 real, 627 fake, labels withheld)
- Audio: 16 kHz, mono, 16-bit PCM WAV, exactly 4 seconds
The test set is speaker-disjoint (no test speaker appears in training) and includes audio from at least one TTS system not seen during training.
Sources
Real speech:
- VCTK Corpus — 109 English speakers, various accents (CC BY 4.0)
- LibriSpeech — read speech from audiobooks (CC BY 4.0)
Fake speech:
- Tacotron2 — attention-based seq2seq TTS (Shen et al., 2018)
- VITS — end-to-end TTS with normalizing flows (Kim et al., 2021)
- SpeechT5 — unified-modal encoder-decoder (Ao et al., 2022)
Files
| File | Description |
|---|---|
train/real/ |
2,874 real speech WAV files |
train/fake/ |
2,874 TTS-generated WAV files |
test/ |
1,254 test WAV files, flat directory, no labels |
train.csv |
Training labels: id, label |
test.csv |
Test file list: id (no label) |
sample_submission.csv |
Example submission with score=0.5 for all |
solution.csv |
Ground truth labels (for local AUROC scoring) |
notebooks/ |
6 reference notebooks (starter + 5 graded approaches) |
train.csv
id,label
train/real/real_00001.wav,real
train/real/real_00002.wav,real
train/fake/fake_00001.wav,fake
train/fake/fake_00002.wav,fake
...
label:"real"or"fake"
test.csv
id
test/voiceguard_00001.wav
test/voiceguard_00002.wav
...
sample_submission.csv
id,score
test/voiceguard_00001.wav,0.5
test/voiceguard_00002.wav,0.5
...
Replace 0.5 with your model's P(fake) score for each clip.
Directory Structure
train/
├── real/
│ ├── real_00001.wav # VCTK / LibriSpeech recordings
│ └── ... # 2,874 files total
└── fake/
├── fake_00001.wav # Tacotron2 / VITS / SpeechT5 output
└── ... # 2,874 files total
test/
└── voiceguard_00001.wav ... voiceguard_01254.wav # flat, no labels
Loading Audio
import librosa
import numpy as np
import pandas as pd
train_df = pd.read_csv("train.csv")
train_df["label_int"] = (train_df["label"] == "fake").astype(int)
y, sr = librosa.load(train_df.iloc[0]["id"], sr=16000)
# y.shape = (64000,) — exactly 4 seconds at 16 kHz
# Anti-spoofing features
flatness = librosa.feature.spectral_flatness(y=y).mean()
mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=20)
zcr = librosa.feature.zero_crossing_rate(y).mean()
# Harmonics-to-Noise Ratio (handle infinity!)
f0, voiced, _ = librosa.pyin(y, fmin=50, fmax=400, sr=sr)
# Note: autocorrelation-based HNR may return np.inf for very periodic signals
# Always apply: np.nan_to_num(hnr, nan=0.0, posinf=0.0, neginf=0.0)
Key Acoustic Differences
Understanding why fake speech is detectable helps you build better features:
| Feature | Real Speech | TTS/Fake Speech |
|---|---|---|
| Spectral flatness | Low (~0.001–0.01) — peaky harmonic spectrum | Higher (~0.05–0.3) — smoother spectrum |
| Pitch micro-variation | Natural jitter (±1–3%) | Machine-perfect, overly smooth F0 |
| HNR | Moderate, with natural noise floor | Very high or infinite (no noise) |
| LFCC smoothness | Frame-to-frame variation | Over-smooth transitions |
| Spectral envelope | Slight roughness from vocal tract | Unnaturally clean |
Important Notes
- All WAV files are exactly 4 seconds (64,000 samples). No padding needed.
- Generalization: the test set includes a TTS system not seen during training. Feature-based models (LFCC, spectral features) generalize better than models that memorize artifacts of specific synthesizers.
- AUROC not accuracy: always output probability scores (0–1), not binary predictions. See the Evaluation page.
- Both real and fake speech are English only in this dataset.
- Dataset is CC BY 4.0. Attribution: VCTK (Yamagishi et al.), LibriSpeech (Panayotov et al.), ASVspoof challenge.