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 models. The dataset was created by AKCIT-Deepfake and was last updated on 2025-11-25.
Use Cases
- Train deepfake speech detection models based on the 459,000 labeled audio samples.
- Benchmark zero-shot TTS model outputs for forensic analysis based on the inclusion of multiple synthetic speech generators.
- Develop multilingual audio deepfake detection systems based on the inclusion of both Brazilian and European Portuguese variants.
Strengths
- Contains 459,000 audio samples, providing a substantial corpus for model training.
- Includes both real and synthetic speech generated by multiple zero-shot TTS models, enabling comparative analysis.
- Covers both Brazilian and European Portuguese variants, supporting research on dialectal differences.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- AKCIT-Deepfake
- Collection Method
- Likely contains audio samples of real speech and synthetic speech generated using multiple zero-shot text-to-speech models.
- Freshness
- Last updated 2025-11-25 22:26:37; freshness should be verified.
- Geography
- Covers Brazilian and European Portuguese language variants.