🗣️ BRSpeech-DF: A Deep Fake Synthetic Speech Dataset for Portuguese

🧩 Description

BRSpeech-DF is the first publicly available dataset for deepfake speech detection in Portuguese, covering both Brazilian and European variants. It contains 459,000 audio samples, including both real and synthetic speech generated using multiple zero-shot text-to-speech (TTS) models.

This dataset aims to foster the development of more robust, inclusive, and multilingual audio deepfake detection systems, especially for underrepresented languages such as Portuguese.

---

📂 Dataset Structure

The dataset is organized in a structure compatible with the audiofolder loader:

BRSpeech-DF/
├── train/
│   ├── bonafide/
│   └── spoof/
├── validation/
│   ├── bonafide/
│   └── spoof/
├── test/
│   ├── bonafide/
│   └── spoof/

Each subfolder represents a class:

Class	Description
bonafide	Real human speech recordings
spoof	Synthetic (deepfake) speech generated via zero-shot TTS models

---

🎛️ Dataset Statistics

Split	Bonafide	Spoof	Total
Train	74,163	370,111	444,274
Validation	1,161	5,795	6,956
Test	1,320	6,587	7,907
Total	76,644	382,493	459,137

---

🧠 Zero-Shot TTS Models Used

Each synthetic utterance was generated using the speaker’s own voice as reference and the original text transcript as input.

Model	Reference	Original Sample Rate
Fish-Speech	Liao et al., 2024	48 kHz
XTTS	Casanova et al., 2024	24 kHz
F5-TTS	Chen et al., 2024	24 kHz
YourTTS	Casanova et al., 2022	16 kHz
ToucanTTS	Lux et al., 2024	24 kHz

All samples were resampled to 24 kHz for consistency.

---

🧮 Benchmark Results (from paper)

Detection model performance on BRSpeech-DF:

Detector	EER (%)	ACC (%)
AASIST	48.12	63.0
SLS-ECAPA	30.67	76.0
SSL Anti-Spoof	35.38	70.2
XLSR-MAMBA	63.03	16.7

Models trained exclusively on English data show a significant performance drop on Portuguese speech, demonstrating the importance of language-specific datasets like BRSpeech-DF.

---

📊 Data Format

Each entry returned by audiofolder includes:

Field	Type	Description
audio	Audio	Path and waveform (24 kHz)
label	ClassLabel	bonafide or spoof

---

📚 Citation

If you use this dataset, please cite:

@inproceedings{BRSpeechDF2025,
  title={BRSpeech-DF: A Deep Fake Synthetic Speech Dataset for Portuguese},
  year={2025},
  note={AKCIT Speech Group}
}

---

⚖️ License

This dataset is distributed under the CC BY 4.0 license. You are free to use it for academic and commercial purposes, provided that proper credit is given.

---

🙌 Credits

• Created by AKCIT-Speech Group
• Real audio derived from LibriVox / Project Gutenberg
• Synthetic audio generated with open-source TTS models (Fish-Speech, XTTS, F5-TTS, YourTTS, ToucanTTS)

---

🔗 Related Resources

• ASVspoof 2021 DF Challenge
• Fish-Speech (Hugging Face)
• Coqui XTTS
• F5-TTS
• YourTTS
• ToucanTTS