ParlaSpoof-BR: A Brazilian Portuguese Political Speech Audio Deepfake Dataset

ParlaSpoof-BR is the first audio deepfake detection benchmark specifically designed for political speech in Brazilian Portuguese. Constructed from official recordings of the Brazilian Chamber of Deputies, the dataset provides a domain-specific evaluation framework for electoral integrity and assesses detector robustness against modern zero-shot voice cloning, voice conversion, and semantically targeted partial manipulation (speech infilling).

---

📌 Key Features

• Authentic Bona Fide Utterances: 2,000 real speech samples from 40 parliamentarians (20 male, 20 female), balanced across all 5 geographic regions of Brazil (North, Northeast, Center-West, Southeast, and South).
• Full-Synthesis Attacks (TTS & VC):
  • 5 Zero-Shot TTS Models: Chatterbox Multilingual V3, XTTS-v2, OmniVoice, VoxCPM2, and Qwen3-TTS.
  • 5 Voice Conversion (VC) Models: Seed-VC, kNN-VC, OpenVoice-v2, X-VC, and EZ-VC.
• Partial Manipulation Attacks (Speech Infilling):
  • Masked infilling via OmniVoice with word alignments provided by WhisperX.
  • LLM-Guided Semantic Attacks: Meaning-inverting edits (antonyms, numbers, names, phrases, and negations) preserving acoustic context.
  • Contiguous Resynthesis: Span resynthesis covering 25%, 50%, and 75% of the utterance duration.
• Acoustic Robustness Perturbations:
  • Babble Noise Injection: Parliamentary background chatter added at 20 dB, 15 dB, and 10 dB SNR.
  • Lossy Codec Compression: Roundtrip transcoding for MP3 and OGG formats.
  • Speech Enhancement: Resemble Enhance, Demucs, and MetricGAN+.
• Total Scale: 134,400 audio files (11,200 bona fide / 123,200 spoofed).

---

📊 Dataset Composition

The benchmark is partitioned into a Core Evaluation Set (unperturbed primary set) and a Robustness Set (acoustic perturbations and codec variants).

Label	Subset / Source	File Count
Bona fide	Original Chamber of Deputies recordings	2,000
Bona fide	Speech enhancement variants (Resemble Enhance, Demucs, MetricGAN+)	6,000
Bona fide	Transcoded variants (MP3 / OGG $\rightarrow$ WAV)	3,200
Spoof	Text-to-Speech (TTS) — 5 generators	10,000
Spoof	Voice Conversion (VC) — 5 systems	10,000
Spoof	Partial Manipulation (OmniVoice Infilling)	8,000
Spoof	Babble Noise Injection (SNRs: 10, 15, 20 dB)	60,000
Spoof	Transcoded variants (MP3 / OGG $\rightarrow$ WAV)	35,200
(Excluded)	Voice-cloning reference prompts ($\ge$ 10s per speaker)	200
Summary	Core Evaluation Set	30,000
Summary	Robustness Variants Set	104,400
TOTAL	Total Benchmark Files	134,400

---

📈 Benchmark Detector Performance

We evaluated three state-of-the-art audio anti-spoofing architectures (AASIST, AASIST-L, and DF-Arena-1B). The empirical results reveal a severe cross-domain generalization gap when models trained on standard benchmarks (such as ASVspoof) are applied to real-world political speech in Brazilian Portuguese.

Overall Performance on Core Evaluation Set (30,000 files)

Detector	EER (%) $\downarrow$	AUC $\uparrow$	Precision	Recall	Macro-F1	Accuracy (%)
AASIST	50.98	0.481	0.911	0.923	0.503	84.85
AASIST-L	53.70	0.445	0.911	0.962	0.500	87.99
DF-Arena-1B	32.30	0.715	0.944	0.830	0.595	80.05

Note: AASIST and AASIST-L exhibit severe bias, classifying over 90% of genuine parliamentary speech as spoofed (False Positive Rates of 91.7% and 95.4%).

DF-Arena-1B Performance by Synthesis Generator

Attack Family	Generator / Method	EER (%) $\downarrow$	AUC $\uparrow$	Recall (%) $\uparrow$
VC	OpenVoice-v2	19.6	0.892	99.7
VC	kNN-VC	21.4	0.849	99.4
VC	X-VC	23.2	0.820	98.1
VC	Seed-VC	30.0	0.743	94.0
VC	EZ-VC	36.0	0.671	83.1
TTS	XTTS-v2	25.5	0.808	95.8
TTS	Chatterbox Multilingual V3	26.1	0.799	96.5
TTS	OmniVoice	33.9	0.705	87.5
TTS	VoxCPM2	53.1	0.463	41.4
TTS	Qwen3-TTS	58.0	0.390	31.2

---

🔍 Key Findings & Bias Analysis

1. Methodological Bias Dominates Demographic Factors:
   • Synthesis Generator Choice: Massive disparity in detectability (a 68.5 pp gap between OpenVoice-v2 and Qwen3-TTS). Portuguese-optimized TTS models (e.g., VoxCPM2, Qwen3-TTS) evade detection at rates exceeding 60%.
   • Gender & Region: Demographic disparities are minimal in comparison (a 0.7 pp gap for gender and a 3.7 pp gap across regional accents).
2. Efficacy of Partial Manipulation:
   • Replacing a few consequential words is significantly more stealthy than full utterance synthesis. The detection rate (recall) for DF-Arena-1B drops to 29.2% on 25% modified audio (compared to 73.5% on 75% modified audio).
3. Codec Asymmetry and Noise Vulnerability:
   • Genuine audio transcoded via OGG causes a catastrophic 94.8% false positive rate in DF-Arena-1B, as compression artifacts closely mimic neural vocoder signatures.
   • Adding authentic background chatter (babble noise at 10 dB SNR) allows 22.2% of previously detected deepfakes to successfully evade detection.

---

📁 File Structure & Metadata

Each sample in the dataset includes automated transcriptions, word-level alignments (WhisperX), and rich metadata annotations:

{
  "id": "parlaspoof_br_001234",
  "speaker_id": "dep_042",
  "gender": "female",
  "region": "Norte",
  "label": "spoof",
  "attack_category": "partial_manipulation",
  "generator_model": "OmniVoice",
  "infill_strategy": "LLM_semantic_antonym",
  "modification_percentage": 0.15,
  "transcription": "O projeto de lei foi rejeitado na comissão.",
  "audio": {
    "path": "audio/spoof/infill/parlaspoof_br_001234.wav",
    "sampling_rate": 16000
  }
}

---

📜 License and Ethical Use

The ParlaSpoof-BR dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0), inheriting the open redistribution license of the official Chamber of Deputies audio archive.

• Intended Use: Academic research, benchmarking audio forensics tools, supporting fact-checking initiatives, and bolstering democratic electoral integrity.
• Prohibited Use: Malicious impersonation, synthesis of political disinformation, or any activity that violates electoral integrity regulations is strictly forbidden.

---