Skip to main content
Data in Brief logoLink to Data in Brief
. 2026 Aug 30;68:113214. doi: 10.1016/j.dib.2026.113214

A dataset of genuine and synthetic Portuguese speech covering European, Brazilian, African, and regional accents - An open resource for speech deepfake detection and trustworthy AI research

André Miguel Viegas Oliveira a,⁎, Rafael Geraldo dos Santos a, Helder Daniel b, José Valente de Oliveira b
PMCID: PMC13571521  PMID: 42733559

Abstract

Portuguese is spoken by approximately 300 million people, making it one of the most widely spoken languages in the world, with speakers distributed across every continent. Despite its global reach, Portuguese remains significantly underrepresented in current synthetic speech detection technologies. This dataset aims to address this critical gap in the existing detection solutions. This dataset may be used in tasks related to biometric security, speaker verification, anti-spoofing systems, and deepfake protection, supporting the development and evaluation of models for Portuguese language speech data.

The Synthetic Speech Detection - Portuguese (SSD_PT) dataset consists of 242,156 audio samples in the Portuguese language, containing short speech segments ranging from 1 to 7 s each. The dataset is organized into two balanced classes: genuine speech samples (bonafide) and synthetic speech samples (spoof).

For the Brazilian Portuguese portion of the dataset, the genuine speech samples were collected from publicly available repositories, namely VoxForge, Mozilla Common Voice, and OpenSLR. As no publicly available repositories with comparable coverage were found for the remaining Portuguese varieties, the genuine speech samples for the remaining portions of the dataset were collected from television interviews, radio programs, and podcasts representing different regions of the Lusophone world, including mainland Portugal, the Azores, Madeira, Angola, and Mozambique.

The synthetic speech samples were generated from genuine speech material. For Brazilian Portuguese, a text-to-speech model based on Coqui TTS (XTTS-v2) was used. For other Portuguese variants, voice conversion techniques were applied using the RVC (Retrieval-based Voice Conversion) framework, requiring the creation of speaker-specific models trained on multiple audio segments to capture individual vocal characteristics.

All audio files were standardized to a 16 kHz sampling rate and converted to mono format. The dataset is split into training, development, and evaluation subsets, accompanied by protocol files that associate each sample with its corresponding label.

Keywords: ASVspoof challenge, Audio deepfakes, Voice conversion, Synthetic speech detection, Portuguese speech


Specifications Table

Subject Computer Sciences
Specific subject area Detection of synthetic Portuguese speech across regional, dialectal and pronunciation variability
Type of data Processed
.wav (audio files), .csv (metadata)
Data collection In a first phase, bonafide (genuine) audio samples were collected from television interviews, radio programs, and podcasts using yt-dlp from DW Notícias, RTP Angola, Rádio Moçambique, RTP Açores, RTP Madeira, RUM – Campus Verbal, RUAFM, and Voz da Planície, as well as from publicly available Brazilian Portuguese speech repositories, namely VoxForge, Mozilla Common Voice, and OpenSLR.
In a second phase, at the University of Algarve (UAlg), within the scope of NOVA LINCS, in Faro, Portugal, spoof (synthetic) audio samples were generated using CoquiTTS (XTTS-v2) and RVC (Retrieval-based Voice Conversion).
Data source location Data are stored at the POLEN DataHub Research Data Repository provided by FCCN, the digital services unit of the FCT – The Portuguese Foundation for Science and Technology, with the aim of promoting the openness and sharing of data generated through funded research, thereby serving the Portuguese scientific community.
The platform is aligned with Open Science practices and the FAIR principles for research data management.
Data accessibility Repository name: POLEN DataHub
Data identification number: https://doi.org/10.57979/UC876E
Direct URL to data: https://datahub.polen.fccn.pt/dataset.xhtml?persistentId=doi:10.57979/UC876E
Related research article none

1. Value of the Data

  • •

    The dataset fills an important gap in the availability of publicly accessible resources for synthetic speech detection in Portuguese by providing a large-scale corpus composed of both bonafide (genuine) and spoof (synthetic) speech samples.

  • •

    In addition to European Portuguese, the dataset includes accents, dialects, and regional variants from mainland Portugal, the Azores, Madeira, Brazil, Angola, and Mozambique. This linguistic diversity enables researchers to investigate the generalization capability of models across different accents and varieties of the Portuguese language.

  • •

    The dataset is organized into standardized training, development and evaluation subsets. This structure facilitates the training and evaluation of machine learning models for synthetic speech detection, speaker verification, and voice anti-spoofing systems.

  • •

    The dataset can be reused to train and evaluate new deep learning architectures, compare different synthetic speech detection approaches, or be extended with additional speech generation techniques.

2. Background

Although several datasets have been developed for voice deepfake detection in English [1], resources available for Portuguese remain limited in terms of both availability and linguistic diversity. This dataset aims to fill this gap, supporting the development and evaluation of synthetic speech detection systems for the Portuguese language, extending the work presented in [2].

This dataset was developed to provide a corpus suitable for training and testing machine learning models designed to distinguish between bonafide human speech and spoofed speech. The data were collected and organized following a binary classification framework widely adopted in automatic speaker verification research and in the ASVspoof challenge [1].

3. Data Description

3.1. Dataset overview

The dataset contains a total of 242,156 audio samples, with durations ranging from 1 to 7 s.

As shown in Table 1, the recordings represent eight Portuguese-language accents and for all practical purposes can be considered balanced, consisting of approximately 50% bonafide samples and 50% spoof samples.

Table 1.

Distribution of genuine and synthetic recordings across the accents included in the dataset.

Accent Genuine Synthetic
Portugal – Azores Islands 16,550 14,652
Portugal – Madeira Islands 18,468 16,931
Portugal – Algarve Region 12,872 15,590
Portugal – Alentejo Region 13,270 15,630
Northern Portugal 14,028 16,880
Mozambique 12,885 14,880
Angola 15,553 16,689
Brazil 14,143 13,135

Although speaker diversity is an important characteristic of speech datasets, the exact number of unique speakers could not be determined for all sources. A substantial portion of the bonafide recordings originated from radio broadcasts, podcasts, and television programmes in which multiple speakers appear, often without visual identification or explicit speaker metadata. Consequently, Table 2 provides a lower-bound estimate of the number of male and female speakers represented in each Portuguese variant. These estimates were obtained through manual inspection of the original media sources whenever speaker identity could be inferred, however, because many recordings contain multiple unidentified speakers, the reported values should be interpreted as conservative estimates of the actual speaker diversity.

Table 2.

Number of unique speakers represented according to Portuguese variant. Values correspond to a lower-bound estimate rather than exact counts.

Accent Bonafide Male Bonafide Female Spoof Male Spoof Female
Portugal – Azores Islands 45 46 7 5
Portugal – Madeira Islands 41 34 5 7
Portugal – Algarve Region 45 36 7 3
Portugal – Alentejo Region 33 27 5 5
Northern Portugal 19 8 6 4
Mozambique 38 10 5 5
Angola 160 80 7 3
Brazil 16 6 13 5

3.2. Dataset split

The dataset is divided into three independent subsets corresponding to the main stages of model development and evaluation, as shown in Table 3.

Table 3.

Dataset split overview.

Subset Directory Percentage Usage in Model Pipeline
Training wav_T 70% Used to train the model
Development wav_D 15% Used at the end of each epoch to monitor performance and compute intermediate metrics
Evaluation wav_E_eval 15% Used only after training completion to compute final model metrics

3.3. File structure

As shown in Fig. 1, the dataset is organized in a hierarchical directory structure. The wav_T, wav_D, and wav_E_eval directories contain only audio files in WAV format, corresponding to the training, development and evaluation subsets, respectively. The SSD_PT_protocols directory stores three TSV protocol files, each associated with one of these dataset partitions. Specifically, SSD_PT.train.tsv, SSD_PT.dev.track_1.tsv, and SSD_PT.eval.track_1.tsv contain the metadata, including the audio filename, attack type and ground-truthlabel, for the training (wav_T), development (wav_D) and evaluation (wav_E_eval) subsets, respectively.

Fig. 1.

Fig 1 dummy alt text

Dataset directory structure.

In addition to the audio and protocol files, the dataset repository also includes a README file and a Data Dictionary. The README provides a brief overview of the dataset, while the Data Dictionary offers a more detailed description of the dataset structure, file naming conventions, and labelling scheme.

3.4. Protocol files description

The protocol files are stored in TSV (Tab-Separated Values) format, however, despite this designation, they are in fact space-separated text files. Each row corresponds to an audio sample and includes the filename, attack type and its corresponding classification label, as illustrated in Table 4.

Table 4.

Example protocol entries.

Filename Attack Label
PT_Algarve_Genuine_20412 bonafide bonafide
PT_Alentejo_Synthetic _126799 RVC spoof

Since the fields in the protocol files are space-separated rather than tab-separated, standard TSV parsers should not be used with the default tab delimiter. For example, when using pandas.read_csv(), the whitespace separator must be explicitly specified to correctly parse the protocol files. The following example illustrates how a protocol file can be loaded using pandas:

import pandas as pd

protocol = pd.read_csv(

 "SSD_PT.train.tsv",

 sep=r"\s+",

 header=None

)

Additional columns are also included and populated with the value “x”, in order to ensure compatibility with formats widely adopted in synthetic speech detection datasets [3]. These values act as placeholders for metadata fields (e.g., speaker ID, gender or codec), which are not considered in this dataset.

3.5. Filename convention

Each audio file follows a structured naming convention, as illustrated in Table 5, which encodes information about geographical origin and speech type.

Table 5.

Describes each component of the filename.

Component Description Example
Region Geographical origin or accent PT_Algarve
DataType Speech type (genuine or synthetic) Genuine
ID Unique sample identifier 20412

The general filename structure is:

[Region]_[DataType]_[ID]

3.6. Attack definition

As shown in Table 6, the Attack column specifies the speech generation method associated with each audio sample.

Table 6.

Defines the possible values of the Attack column.

Attack Description
bonafide Speech produced by a human speaker.
RVC Speech generated using a Retrieval-based Voice Conversion (RVC) system.
CoquiTTS Speech generated using the CoquiTTS (XTTS-v2) system.

3.7. Label definition

The dataset follows a binary classification scheme consisting of two labels, as illustrated in Table 7: bonafide and spoof.

Table 7.

Label definitions.

Label Description
bonafide Speech produced by a human speaker.
spoof Synthetic or manipulated speech generated using TTS or voice conversion systems.

3.8. Benchmark protocol

To ensure reproducibility and enable a fair comparison between future synthetic speech detection methods, the SSD_PT dataset provides predefined training, development, and evaluation splits, as previously described. Models should be trained exclusively on the training set, while the development set is used to monitor the training process and select the hyperparameters. The final performance should be reported solely based on the evaluation set.

The recommended evaluation metrics are Accuracy, Sensitivity, Specificity, and Equal Error Rate (EER). True Positives (TP) denote spoof samples correctly classified as spoof; True Negatives (TN) denote bonafide samples correctly classified as bonafide; False Positives (FP) denote bonafide samples incorrectly classified as spoof; and False Negatives (FN) denote spoof samples incorrectly classified as bonafide. Accuracy, Sensitivity, and Specificity are computed according to Eqs. (1)–(3).The evaluation metrics are computed as follows:

Accuracy=TP+TNTP+TN+FP+FN (1)
Sensitivity=TPTP+FN (2)
Specificity=TNTN+FP (3)

The Equal Error Rate (EER) corresponds to the operating point at which the False Acceptance Rate (FAR) equals the False Rejection Rate (FRR).

4. Experimental Design, Materials and Methods

4.1. Bonafide audio collection

The bonafide speech recordings were collected from publicly available sources representing different Portuguese-speaking regions. For the European Portuguese, Angolan Portuguese, and Mozambican Portuguese accents, recordings were obtained from television broadcasts, radio programmes, interviews, and podcasts. Specifically, audio data were collected from RTP Açores [4] for the Azores accent, RTP Madeira [5] for the Madeira accent, RUA FM [6] for the Algarve accent, RUM – Campus Verbal [7] for the Northern Portugal accent, Voz da Planície [8] for the Alentejo accent, Rádio Moçambique [9] for the Mozambican accent, and DW Notícias [10] together with RTP Angola [11] for the Angolan accent.

Audio extraction from these sources was performed using the yt-dlp framework [12], an open-source command-line tool widely used for downloading audio and video content from online platforms.

After audio collection, given that the recordings were lengthy, voice activity detection (VAD) was applied to segment the audio into shorter excerpts, removing prolonged silences and noise. Subsequently, these segments were randomly selected, ensuring speaker and content diversity while avoiding continuous reproduction of the original recordings.

For Brazilian Portuguese, bonafide speech recordings were obtained from publicly available speech repositories, namely VoxForge [13], Mozilla Common Voice [14] and OpenSLR [15]. These repositories provide speech recordings produced by native speakers.

Finally, all audio samples were converted from various formats into uncompressed WAV files, followed by resampling to 16 kHz and conversion to mono channel. The resulting recordings were subsequently incorporated into the dataset as bonafide samples.

The source recordings were obtained exclusively from publicly accessible broadcasts and repositories whose terms permit research use or whose use is authorised under applicable copyright exceptions for scientific research.

4.2. Spoof audio generation

Synthetic speech samples were generated from the bonafide recordings using speech synthesis and voice conversion frameworks. Different approaches were adopted.

4.3. Brazilian Portuguese

For Brazilian Portuguese, spoof samples were generated using CoquiTTS (XTTS-v2) [16]. XTTS-v2 is a multilingual text-to-speech framework capable of generating synthetic speech while preserving speaker-specific vocal characteristics from a reference audio sample.

The generation process required:

  • 1.

    A reference speech sample from the target speaker;

  • 2.

    A text transcription;

  • 3.

    The pre-trained XTTS-v2 model.

No additional model training was performed. Synthetic audio was directly generated using the pre-trained model and the provided speaker reference.

European Portuguese, Angolan Portuguese, and Mozambican Portuguese

Since XTTS-v2 does not currently provide native support for European Portuguese, synthetic speech for the remaining accents was generated using RVC (Retrieval-based Voice Conversion) [17].

For each accent, between 10 and 12 speakers were selected. For each speaker, approximately one to two hours of speech recordings were collected and used to train an independent RVC model. During the training process, each model learned the vocal characteristics of the corresponding speaker from the available data.

After the training phase, genuine speech recordings were used as input to the trained models. The RVC framework generated new synthetic audio samples, preserving the original linguistic content of the input recordings while converting the vocal identity to the voice of the target speaker represented by the trained model. On average, between approximately 1200 and 1700 synthetic audio samples were generated per model.

All audio files were converted to a 16 kHz sampling rate and mono format. The resulting synthetic recordings were subsequently incorporated into the dataset as spoof samples.

The full audio processing pipeline was implemented using custom Python scripts developed for this study. These scripts cover audio extraction, VAD-based segmentation, and preprocessing steps applied to both bonafide and spoof data. The complete source code is publicly available from a GitHub repository, archived on Zenodo to ensure reproducibility [18].

The audio generative models were trained exclusively for the purpose of generating synthetic speech samples used in this dataset. The trained voice-conversion models are not distributed as part of the released dataset. Only the resulting synthetic audio samples are made publicly available, together with the scripts used to construct the dataset, thereby ensuring reproducibility while preventing redistribution of the speaker-specific voice-conversion models.

Limitations

A limitation of the dataset is that it relies exclusively on two spoof audio generative methods, RVC and CoquiTTS (XTTS-v2), which may restrict the diversity and generalizability of spoofing scenarios represented. Future extensions could incorporate additional speech synthesis and voice conversion architectures, such as FreeVC [19] and YourTTS [20], to broaden the range of synthetic speech generation techniques and spoofing scenarios covered by the dataset.

CRediT Author Statement

André Miguel Viegas Oliveira: Methodology, Writing - Original Draft, Writing - Review & Editing, Data Curation, Visualization. Rafael Geraldo dos Santos: Methodology, Writing - Review & Editing. Helder Daniel: Methodology, Writing - Review & Editing, Resources. José Valente de Oliveira: Methodology, Writing - Review & Editing, Project administration.

Ethics Statement

The dataset is composed of audio segments of up to 7 s of duration collected from publicly available audio and audiovisual sources. The audio segments were randomly clipped from long recordings, stripped of all identifying metadata, thoroughly shuffled across sources, assigned unrelated filenames, before their inclusion in the dataset. Those audios collected from VoxForge, Mozilla Common Voice, and OpenSLR do not require any usage consent. All the others, including accessible radio broadcasts, and online media comply with the applicable copyright exceptions for scientific research and educational purposes under the legal frameworks of Portugal (Decree-Law No. 47/2023, amending Decree-Law No. 63/85), Angola (Law No. 15/14), and Mozambique (Law No. 9/2022).

Moreover, besides the above mentioned audio clips, no other person identifiable or otherwise data were collected.

Because all source recordings were obtained from publicly available repositories or lawfully accessible public broadcasts and were used exclusively for scientific research under the applicable legal frameworks, explicit individual consent from the speakers was not required under the applicable legal framework governing the use of publicly available recordings for scientific research.

The trained generative models are not distributed, are not reusable for speaker impersonation, and were employed exclusively as intermediate research tools to generate synthetic speech samples for machine-learning research. The released dataset contains only short speech excerpts and synthetic outputs. It is not intended nor suitable for recreating the vocal identity of any individual speaker.

Acknowledgements

Funding: This work was supported by NOVA LINCS (UID/04516/2025; DOI: 10.54499/UID/04516/2025) and by the project 2024.07291.IACDC – PT-SSD – Portuguese Synthetic Speech Detection for Public Administration, funded by FCT (The Portuguese Foundation for Science and Technology).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References


Articles from Data in Brief are provided here courtesy of Elsevier

RESOURCES