Synthetic speech detection in Brazilian Portuguese through accent-related features

Authors: Pedro H. L. Leite, Pedro Benevenuto Valadares, Luiz Wagner Pereira Biscainho

Published: 2026-09-20 18:52:40+00:00

Comment: \\c{opyright} 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works

AI Summary

This work introduces a speech deepfake detection methodology for Brazilian Portuguese (pt-BR) by exploiting the dialectal inconsistencies of synthetic speech. It combines multilingual phone recognizers and classical signal processing to extract phoneme-level features with high geographic variance. The approach reveals that the distributional gap in these accent-related features effectively distinguishes natural and synthetic voices, establishing dialectal inconsistency as an interpretable cue for spoofing detection.

Abstract

Leading commercial and open-source Text-to-Speech (TTS) models fail to emulate the regional phonetic diversity of Brazilian Portuguese (pt-BR). By aggregating disparate dialects into a single training distribution, they generate a synthetic diluted accent: a phonetic profile attempting to represent all regional distributions simultaneously, but ultimately carrying phonological ambiguity dissociated from natural socio-phonetic realizations. This work introduces a speech deepfake detection methodology combining multilingual phone recognizers with classical signal processing to extract phoneme-level features in consonantal and vocalic realizations with high geographic variance. The analysis reveals that the distributional gap over these features suffices to distinguish natural and synthetic voices through unsupervised Kernel Density Estimation, establishing dialectal inconsistency as a useful and interpretable feature for spoofing detection in pt-BR. Evaluation on pt-BR anti-spoofing datasets shows that these explainable, lightweight, low-dimensional features can boost the performance of foundation models on the task, and show generalization capabilities in a cross-dataset leave-one-out setup.

Key findings

The distributional gap in accent-related features between natural and synthetic speech is sufficient for unsupervised detection, with KDE achieving 74.3% AUC (93.9% after t-SNE). Synthetic voices exhibit interpretable shifts in distributions, distorted spectral shapes, and higher within-utterance marker switching. These explainable, lightweight features enhance the performance of foundation models in deepfake detection, achieving 2.00% EER on BRSpeechDF and demonstrating strong cross-dataset generalization (e.g., 6.09% overall EER in a double leave-one-dataset-out setup with ECAPA+Formants).

Approach

The authors propose extracting phoneme-level features from consonantal and vocalic realizations, which exhibit high geographic variance in Brazilian Portuguese. They use multilingual phone recognizers (ZIPA, PhoneticXeus) and classical signal processing to create speaker-level phonetic profiles. These profiles are then used with unsupervised Kernel Density Estimation and various supervised classifiers to detect synthetic speech based on dialectal inconsistencies.

Datasets

Custom curated dataset (364 natural speakers from CORAA, Mozilla Common Voice, ColingPB, CML-TTS, NURC-SP, NURC-RE, Certas Palavras, CETUC, Tagarela; 57 synthetic voices from Azure, Google, OpenAI, ElevenLabs, F5-TTS, Qwen3-TTS, Piper, Kokoro), BRSpeechDF, FakeBRAccent, MLAAD (Portuguese subset).

Model(s)

Multilingual Phone Recognizers (ZIPA, PhoneticXeus), Supervised Classifiers (XGBoost, Random Forest, SVM, Logistic Regression), Foundation Models (Wav2Vec2Bert, XLSR, ECAPA-TDNN, HuBERT), Signal Processing (Parselmouth for formant extraction).

Author countries

Brazil

← Previous