Upload dataset.md with huggingface_hub

dataset.md

@@ -0,0 +1,171 @@
+# VoiceGuard — Dataset Documentation
+
+## Overview
+
+**Task:** Deepfake Audio Detection — predict the probability that a 4-second audio clip is AI-generated (fake) vs. real human speech.  
+**Metric:** AUROC (Area Under ROC Curve)  
+**HuggingFace:** `fassabilf/voiceguard-competition`
+
+---
+
+## Sources
+
+**Real speech:**
+- [VCTK Corpus](https://datashare.ed.ac.uk/handle/10283/3443) — 109 speakers, various English accents (CC BY 4.0)
+- [LibriSpeech](https://www.openslr.org/12) — read English speech from audiobooks (CC BY 4.0)
+
+**Fake speech (AI-generated):**
+- Generated by ≥3 TTS systems: **Tacotron2**, **VITS**, **SpeechT5**
+- Test set includes ≥1 unseen TTS system (not present in training)
+
+---
+
+## Statistics
+
+| Split | Real | Fake | Total |
+|-------|------|------|-------|
+| Train | 2,874 | 2,874 | 5,748 |
+| Test  | 627  | 627  | 1,254 |
+
+- **Class balance:** perfectly balanced (50% real, 50% fake)
+- **Baseline (random score 0.5):** AUROC = 0.5
+- **Speaker-disjoint:** test speakers ≠ train speakers
+
+---
+
+## Audio Format
+
+| Property | Value |
+|----------|-------|
+| Sample rate | 16,000 Hz |
+| Channels | Mono |
+| Duration | 4 seconds (padded/trimmed) |
+| Format | WAV (PCM 16-bit) |
+
+---
+
+## File Structure
+
+```
+voiceguard/
+├── train/
+│   ├── real/
+│   │   ├── real_00001.wav
+│   │   └── ...           # 2,874 files
+│   └── fake/
+│       ├── fake_00001.wav
+│       └── ...           # 2,874 files
+├── test/
+│   ├── voiceguard_00001.wav
+│   └── ...               # 1,254 files, flat (no labels)
+├── train.csv
+├── test.csv
+├── sample_submission.csv # id, score=0.5
+└── solution.csv          # id, label (real/fake)
+```
+
+---
+
+## CSV Columns
+
+**train.csv**
+```csv
+id,label
+train/real/real_00001.wav,real
+train/fake/fake_00001.wav,fake
+...
+```
+
+**sample_submission.csv / your submission**
+```csv
+id,score
+test/voiceguard_00001.wav,0.92
+test/voiceguard_00002.wav,0.04
+...
+```
+
+`score` = P(fake) — probability between 0 and 1.  
+Higher score = more likely to be fake.  
+**Do NOT submit hard labels (0/1) — submit probabilities.**
+
+---
+
+## Split Construction
+
+1. Real clips sampled from VCTK + LibriSpeech, trimmed to 4 seconds
+2. Fake clips generated via Coqui TTS (Tacotron2, VITS) and Microsoft SpeechT5
+3. Stratified 80/20 split, **speaker-disjoint**: no speaker appears in both train and test
+4. At least 1 TTS system in test not seen during training (generalization challenge)
+5. Test files renamed to flat `voiceguard_NNNNN.wav`
+
+---
+
+## Loading the Data
+
+```python
+import pandas as pd
+import librosa
+import numpy as np
+
+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)
+label = train_df.iloc[0]["label"]       # 'real' or 'fake'
+label_int = train_df.iloc[0]["label_int"]  # 0 or 1
+
+# Spectral flatness (key anti-spoofing feature)
+flatness = librosa.feature.spectral_flatness(y=y).mean()
+print(f"Spectral flatness: {flatness:.4f}")
+# Real speech: ~0.001-0.01 (peaky spectrum)
+# TTS speech: ~0.05-0.3  (flatter spectrum)
+```
+
+---
+
+## Key Discriminative Features
+
+TTS systems leave detectable artifacts that distinguish fake from real:
+
+| Feature | Real Speech | TTS/Fake Speech |
+|---------|-------------|-----------------|
+| Spectral flatness | Low (~0.001–0.01) | High (~0.05–0.3) |
+| HNR (Harmonics-to-Noise) | Moderate variability | Very high / uniform |
+| LFCC smoothness | Natural variation | Over-smooth |
+| Pitch trajectory | Natural micro-variation | Machine-perfect |
+| Spectral envelope | Slight roughness | Clean, artifact-free |
+
+These differences make this dataset **easily separable** (AUROC=1.0 on all approaches).
+
+---
+
+## Evaluation
+
+Score = AUROC on test set (threshold-free ranking metric):
+```python
+from sklearn.metrics import roc_auc_score
+score = roc_auc_score(y_true_binary, y_pred_scores)
+# y_true_binary: 1=fake, 0=real
+# y_pred_scores: your P(fake) predictions
+```
+
+AUROC = 1.0 means perfect ranking; AUROC = 0.5 means random.
+
+---
+
+## Difficulty & Insights
+
+- **Why AUROC=1.0?** This dataset uses modern but detectable TTS systems. Real speech from VCTK/LibriSpeech has natural vocal tract noise; Tacotron2/VITS produce unrealistically clean spectral envelopes.
+- **Production hardness:** Real-world deepfake detection is much harder. See [ASVspoof 2021 DF](https://www.asvspoof.org/) and [In-the-Wild](https://deepfake-total.com/) for challenging benchmarks.
+- **Unseen TTS generalization:** The held-out TTS system in the test set is included to test generalization. Feature-based approaches (LFCC, HNR) generalize better than systems overfitting to one TTS artifact.
+
+---
+
+## Real-World Context
+
+Voice deepfakes are increasingly used in:
+- **Voice phishing (vishing):** impersonating executives or family members
+- **Identity fraud:** bypassing voice-based authentication
+- **Disinformation:** fabricated statements attributed to public figures
+
+Key benchmarks beyond this dataset: [ASVspoof 2019/2021](https://www.asvspoof.org/), [ADD Challenge](https://addchallenge.cn/), [In-the-Wild Dataset](https://deepfake-total.com/).