ParlaSpoof-BR

Audio Deepfake Detection Benchmark for Brazilian Political Speech

134,400 samples | 40 speakers | 5 regions

Dataset Overview

ParlaSpoof-BR is the first public audio deepfake benchmark targeting Brazilian Portuguese political speech. The dataset is derived from recordings of the Brazilian Chamber of Deputies and includes realistic synthetic utterances generated using state-of-the-art text-to-speech and voice conversion models.

134,400
Total Samples
30,000
Core Evaluation Set
40
Speakers
5
Regions
Label Subset Files
Bona Fide Original Chamber recordings 2,000
Enhanced (Resemble Enhance, Demucs, MetricGAN+) 6,000
MP3/OGG → WAV variants 3,200
Spoof TTS: OmniVoice, XTTS-v2, Chatterbox, VoxCPM2, Qwen3-TTS 10,000
Voice Conversion: Seed-VC, kNN-VC, OpenVoice-v2, XVC, EzVC 10,000
OmniVoice speech infilling (25%, 50%, 75%, LLM Attack) 8,000
Babble noise: three SNR levels (10/15/20 dB) 60,000
MP3/OGG → WAV variants 35,200

Audio Samples

Text-to-Speech (TTS) Comparison

TTS models synthesize speech from text using a reference voice. Below we show: (1) the source text audio from Beto Pereira, (2) the reference voice from Abilio Brunini, and (3) synthesized outputs from 5 TTS models cloning Abilio's voice to speak Beto's text.

Reference Inputs

SOURCE TEXT

Beto Pereira (Original Speaker)

The text content to be synthesized

REFERENCE VOICE

Abilio Brunini (Target Voice)

Voice characteristics to clone

Synthesized Outputs (5 TTS Models)

TTS

OmniVoice

MaskGIT-based architecture with masked infilling

TTS

XTTS-v2

Multilingual (17 languages) zero-shot cloning

TTS

Chatterbox

Zero-shot voice cloning

TTS

VoxCPM2

Portuguese-optimized synthesis

TTS

Qwen3-TTS

1.7B parameter multilingual model

Voice Conversion (VC) Samples

VC models transform speaker identity while preserving linguistic content. Below we show: (1) the source audio from Beto Pereira, (2) the reference voice from Abilio Brunini, and (3) converted outputs from 5 VC models transforming Beto's speech to sound like Abilio.

Reference Inputs

SOURCE AUDIO

Beto Pereira (Original Speech)

Audio to be converted

REFERENCE VOICE

Abilio Brunini (Target Voice)

Voice characteristics to clone

Converted Outputs (5 VC Models)

VC

Seed-VC

Self-supervised voice conversion

VC

kNN-VC

k-Nearest Neighbors voice conversion

VC

OpenVoice-v2

Zero-shot cross-lingual voice cloning

VC

XVC

Expressive voice conversion

VC

EzVC

Easy voice conversion

LLM-Guided Semantic Attacks

An LLM (Claude Sonnet 4) identifies semantically critical words to manipulate, selecting targets across categories: NUMBER, ANTONYM, NAME, PHRASE, and NEGATION. OmniVoice's native masked infilling then regenerates only the target segment, conditioned on surrounding acoustic context.

Example 1: ANTONYM Attack

ANTONYM

Original: "...diretamente tambem podemos antever no horizonte uma crise economica no Estado com reflexo nacional."

Manipulated: "...diretamente tambem podemos antever no horizonte uma prosperidade economica no Estado com reflexo nacional."

Impact: Transforms negative prediction into positive scenario, completely changing the message

ORIGINAL
ATTACKED

Example 2: NEGATION Attack

NEGATION

Original: "E se estamos diante de uma crise financeira e de produtividade no campo..."

Manipulated: "E se nao estamos diante de uma crise financeira e de produtividade no campo..."

Impact: Completely denies the existence of the crisis, inverting the political diagnosis

ORIGINAL
ATTACKED

Example 3: PHRASE Attack

PHRASE

Original: "...venho solicitar ao governo federal que sejam implementadas medidas emergenciais em relacao aos financiamentos agricolas."

Manipulated: "...venho solicitar ao governo federal que sejam implementadas cortes imediatos em relacao aos financiamentos agricolas."

Impact: Transforms request for help into request for resource cuts

ORIGINAL
ATTACKED

Example 4: ANTONYM Attack

ANTONYM

Original: "...venho solicitar ao governo federal que sejam implementadas medidas emergenciais..."

Manipulated: "...venho solicitar ao governo federal que sejam suspensas medidas emergenciais..."

Impact: Completely inverts the request from action to inaction

ORIGINAL
ATTACKED

Example 5: ANTONYM Attack

ANTONYM

Original: "...eu tenho certeza que nos poderemos ter avancos significativos nas pautas de energia."

Manipulated: "...eu tenho certeza que nos poderemos ter retrocessos significativos nas pautas de energia."

Impact: Completely inverts the positive expectation to negative about agenda progress

ORIGINAL
ATTACKED

Example 6: NAME Attack

NAME

Original: "Mas quero aqui tambem dizer aos colegas, como Mato Grossense aqui..."

Manipulated: "Mas quero aqui tambem dizer aos colegas, como Paulista aqui..."

Impact: Changes the declared geographic origin, altering regional context

ORIGINAL
ATTACKED

Example 7: ANTONYM Attack (Historical Falsification)

ANTONYM

Original: "...ele que foi autor da emenda das diretas ja no nosso pais, que proporcionou as eleicoes diretas para presidente..."

Manipulated: "...ele que foi autor da emenda das diretas ja no nosso pais, que impediu as eleicoes diretas para presidente..."

Impact: Changes historical role from facilitator to obstructor of direct elections

ORIGINAL
ATTACKED

Example 8: ANTONYM Attack

ANTONYM

Original: "...entao eu tenho muito orgulho de como matogrossense homenagear esse cidadao..."

Manipulated: "...entao eu tenho muito vergonha de como matogrossense homenagear esse cidadao..."

Impact: Transforms praise into severe criticism of the honoree

ORIGINAL
ATTACKED

Contiguous Resynthesis

Resynthesizes contiguous segments at varying percentages (25%, 50%, 75%) without altering semantic content. This isolates the purely acoustic detectability of partial synthesis. Each row shows one utterance with original and manipulated versions.

Sample 1

ORIGINAL

Unmodified

25%

Minimal

50%

Medium

75%

Extensive

Sample 2

ORIGINAL

Unmodified

25%

Minimal

50%

Medium

75%

Extensive

Babble Noise Robustness

Synthetic audio (OmniVoice TTS) with babble noise at different SNR levels. Lower SNR means more noise, which can help deepfakes evade detection. At 10dB SNR, DF-Arena shows 22.2% evasion rate.

CLEAN TTS

No noise added

Original synthetic audio

SNR 20dB

Light noise

Minimal background chatter

SNR 15dB

Moderate noise

Noticeable crowd noise

SNR 10dB

Heavy noise

Significant masking - 22.2% evasion

Key Findings

32%
DF-Arena-1B EER on ParlaSpoof-BR
71%
Evasion rate for 25% partial manipulation
68.5pp
Detection gap between best/worst TTS models
<4pp
Gender and regional bias

Main conclusion: State-of-the-art detectors are fundamentally unprepared for Brazilian Portuguese political speech. Attack methodology (synthesis model choice, manipulation percentage) dominates over demographics (gender, region) in determining detection success.