Posts
Projects
Publications
About
Brazilian Portuguese Speech Datasets | Lucas H. Ueda

Brazilian Portuguese Speech Datasets

Demo

📅 Modified at: October 26, 2026
Added WorldSpeech (1,763.5h pt-BR) to ASR and Corpus CEFALA-1 (104 speakers) to ASV resources. Adjusted chronological ordering for consistency.

📅 Modified at: September 16, 2026
Added YODAS-Granary (13.4k hours) and the CoLingPB sociolinguistic corpus to the ASR resources.

Here is a comprehensive compilation of speech datasets available for Brazilian Portuguese (pt-BR) categorized by task and ordered chronologically.

Automatic Speech Recognition (ASR)

CoLingPB — Corpus Linguístico da Paraíba (2015)

Unknown size · Raw conversational audio · Multi-turn annotations · Paraíba dialect

A sociolinguistic documentary corpus aimed at preserving the dialectal varieties of the Brazilian state of Paraíba. Recorded between 2011 and 2014, the dataset provides raw audio interviews accompanied by text transcriptions and multi-turn conversational markers. Because the audio files and transcriptions are distributed separately across individual project publications rather than a centralized repository, consolidating the full corpus requires web scraping. The total duration and exact number of utterances are unspecified.

  • Authors: Cirineu Cecote Stein (coordinator), Anna Elizabeth de C. C. da Silveira, Elioenai Macena de Araújo, Felipe de Castro Cruz, Lucas Hudson Pequeno da Silva, Maria Cristina da Silva Lima, Solange Soares Costa
  • Download: Distributed across multiple publications (requires web scraping to consolidate)
  • Source: Funded by MEC/SESu (PROEXT) and executed at UFPB (2011–2015)

Multilingual LibriSpeech (MLS) — pt-BR (2020)

Multi-speaker · Large-scale · Read audiobook · 8 languages

Large multilingual corpus derived from LibriVox audiobooks across 8 languages. The Portuguese partition offers hundreds of hours of read audiobook speech, making it one of the largest publicly available pt-BR ASR datasets.

Sidney & VoxForge (2020)

72 / 111+ speakers · Non-controlled · Diverse tasks · Word-level transcription

Two corpora available from a unified page.

Sidney: 72 speakers (20 women, ages 17–59), 5,777 utterances covering digits, words, phonetic sentences and more, recorded at 22 kHz in non-controlled environments. Fields include place of birth, age, gender, education, and occupation.

VoxForge: Crowd-sourced GPL corpus with 111+ speakers, 4,130 pt-BR utterances at varying sample rates (16–44.1 kHz); highly heterogeneous quality with low SNR on many recordings.

  • Authors: No paper related
  • Download: Datasets Page (contains individual download links for both corpora)

mTEDx (2021)

Multi-speaker · Semi-spontaneous · 8 languages · Sentence-aligned

Sentence-level aligned TEDx Talk recordings and transcripts in 8 languages including pt-BR, with translations into up to 5 languages. Covers natural, semi-spontaneous public speaking across diverse topics.

  • Authors: Elizabeth Salesky, Matthew Wiesner, Jacob Bremerman, Roldano Cattoni, Matteo Negri, Marco Turchi, Douglas W. Oard, Matt Post
  • Download: OpenSLR Repository
  • Paper: ISCA Archive

CORAA v1.1 (2021)

Multi-speaker · ~291 h · 400k+ clips · Varied sources

Large publicly available pt-BR ASR dataset with 290.77 hours of audio and 400k+ segmented transcribed utterances. Covers a broad range of sources and speaking styles.

  • Authors: Arnaldo Candido Junior, Edresson Casanova, Anderson Soares, Frederico Santos de Oliveira, Lucas Oliveira, Ricardo Corso Fernandes Junior, Daniel Peixoto Pinto da Silva, Fernando Gorgulho Fayet, Bruno Baldissera Carlotto, Lucas Rafael Stefanel Gris, Sandra Maria Aluísio
  • Download: Hugging Face
  • Paper: Springer

CORAA NURC-SP (2024)

Multi-speaker · ~240 h · 170k+ clips · Naturalistic · São Paulo dialect

Pt-BR ASR corpus with 239.68 hours (239.30 h filtered) and 170k+ segmented utterances, derived from the NURC-SP sociolinguistic corpus of São Paulo speech. Covers naturalistic spoken language with diverse sociolinguistic profiles.

  • Authors: Rodrigo Lima, Sidney E. Leal, Arnaldo Candido Junior, Sandra M. Aluísio
  • Download: Hugging Face
  • Paper: SBC Open Lib

CORAA MUPE ASR (2025)

289 speakers · 365 h · Spontaneous interviews · Regional diversity

Life story interview corpus with 289 interviews totalling 365 hours. Highly diverse in age, education, and regional accent — well-suited for robustness testing across sociolinguistic variation.

  • Authors: Sidney Evaldo Leal, Arnaldo Candido Junior, Ricardo Marcacini, Edresson Casanova, Odilon Gonçalves, Anderson Silva Soares, Rodrigo Freitas Lima, Lucas Rafael Stefanel Gris, Sandra Aluísio
  • Download: Hugging Face
  • Paper: ACL Anthology

YODAS-Granary (2025)

Multi-speaker · 13.4k h (pt) · Pseudo-labeled · ASR + AST

A curated subset of the larger nvidia/Granary dataset, derived from the espnet/yodas2 corpus. The Portuguese partition contains approximately 13.4k hours of speech. Although officially categorized under European Portuguese in the metadata, the vast majority of the audio features Brazilian Portuguese. It provides high-quality pseudo-labeled data for both Automatic Speech Recognition (ASR) and Automatic Speech Translation (AST). ASR transcriptions were generated using Systran/faster-whisper-large-v3 and refined for punctuation/capitalization with Qwen/Qwen2.5-7B-Instruct, while AST translations to English were generated using utter-project/EuroLLM-9B-Instruct.

  • Authors: Nithin Rao Koluguri, Monica Sekoyan, George Zelenfroynd, Sasha Meister, Shuoyang Ding, Sofia Kostandian, He Huang, Nikolay Karpov, Jagadeesh Balam, Vitaly Lavrukhin, Yifan Peng, Sara Papi, Marco Gaido, Alessio Brutti, Boris Ginsburg
  • Download: Hugging Face
  • Paper: arXiv:2505.13404

TAGARELA (2026)

Multi-speaker · 8,972+ h total · ~2k h TTS subset · Podcast / spontaneous · ASR + TTS

Massive podcast-derived corpus with 8,972+ hours of naturalistic conversational pt-BR. Includes a high-quality ~2k-hour subset curated for TTS. One of the largest open pt-BR speech resources available.

  • Authors: Frederico Santos De Oliveira, Lucas Rafael Stefanel Gris, Alef Iury Siqueira Ferreira, Augusto Seben Da Rosa, Alexandre Costa Ferro Filho, Edresson Casanova
  • Download: Hugging Face
  • Paper: IEEE Xplore

WorldSpeech (2026)

Multilingual · >65k h total · 1,763.5 h pt-BR · 24 kHz

A multilingual ASR dataset containing over 65k hours of human transcribed speech across 127 language-region variants, drawn from national parliaments, public broadcasters, public-domain audiobooks, and international institutions. Rows consist of 24 kHz speech utterances paired with a human-provided transcript, an aligned ASR transcript, character error rate (CER) between the two, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores. Mostly for ASR tasks, it provides 1,763.5 hours of pt-BR speech, sourced primarily from the Brazilian Federal Senate and the Chamber of Deputies (Câmara).

  • Authors: Antonis Asonitis, Luca A. Lanzendörfer, Frédéric Berdoz, Roger Wattenhofer
  • Download: Hugging Face
  • Paper: arXiv

Text-to-Speech (TTS)

Accent Dataset SP-RJ

Dual-accent · Accent conversion · São Paulo & Rio de Janeiro dialects

A specialized dataset designed specifically for accent conversion tasks, featuring speech data from two prominent Brazilian regional dialects: São Paulo and Rio de Janeiro.

  • Authors: Fernanda Lustosa
  • Download: Kaggle
  • Paper: No paper related

TTS Portuguese Corpus (2020)

Single speaker · ~10.5 h · Read speech · 48 kHz

Single-speaker corpus with ~10.5 hours of clean studio-quality speech (48 kHz), comprising 3,632 Wave files ranging from 0.7 to 50 seconds. Designed for TTS but widely used for ASR benchmarking.

  • Authors: Edresson Casanova, Arnaldo Candido Junior, Christopher Shulby, Frederico Santos de Oliveira, João Paulo Teixeira, Moacir Antonelli Ponti, Sandra Maria Aluisio
  • Download: GitHub Repository
  • Paper: Springer

Globo Speech Datasets (2023)

Single speaker each · ~20 h each · Studio quality

High-quality studio-grade speech corpora from Globo, each ~20 hours. Comparable to LJSpeech in quality. Two releases: a female speaker corpus and a general (male) speaker corpus.

CML-TTS (2023)

Multi-speaker · Read audiobook · 7 languages · TTS-optimized

Derived from MLS and adapted for TTS training across 7 languages including pt-BR. Provides cleaner segmentation and metadata more suitable for speech synthesis than the original MLS splits.

  • Authors: Oliveira, Frederico S. and Casanova, Edresson and Junior, Arnaldo Candido and Soares, Anderson S. and Galvão Filho, Arlindo R.
  • Download: GitHub Repository
  • Paper: ACM DL

CommonVoice Consolidated (pt-BR) (2025)

Multi-speaker · Crowd-sourced · TTS-optimized

Preprocessed and merged version of Mozilla Common Voice pt-BR, with filtered, normalized audio clips optimized for both ASR and TTS training pipelines.

FalaBrasil Consolidado (2025)

Multi-speaker · Varied sources · TTS-optimized

Consolidated multi-speaker pt-BR speech corpus assembled from the FalaBrasil fb-audio-corpora GitLab repository, covering a range of speaking conditions suitable for TTS training.

NURC ENTOA TTS (2025)

Multi-speaker · TTS-optimized · 4 configurations · São Paulo dialect

TTS-focused corpus derived from the NURC-SP project with 4 configurations: prosodic, automatic, audioCorpus, and test. Captures naturalistic São Paulo speech adapted for synthesis tasks.

NURC TTS (2026)

Multi-speaker · TTS-optimized · São Paulo + Recife dialects

Two-partition corpus covering São Paulo and Recife dialects. Overlaps with NURC-SP and CORAA datasets. Useful for multi-dialect TTS coverage across Brazilian Portuguese regional varieties.

Speech Emotion Recognition (SER)

VERBO (2018)

Acted · 6 emotions (Ekman) · 12 actors · 1,167 sentences · Discrete labels

Acted emotion database with 1,167 sentences recorded by 12 professional actors. Covers 6 basic emotions from Ekman’s model (happiness, sadness, anger, fear, disgust, surprise). Uses a discrete classification scheme.

emoUERJ (2021)

Acted · 4 emotions · 377 clips · Balanced classes

Small emotion corpus with 377 audio clips across 4 categories: happiness (91), anger (94), sadness (100), and neutral (92). Balanced per-class distribution.

  • Authors: No paper related
  • Download: Zenodo

CORAA SER (2022)

Natural speech · 3 classes · ~50 min · Gender-aware labels

~50 minutes of labeled audio segments in 3 classes: neutral, non-neutral female, and non-neutral male. Derived from naturalistic speech rather than acted performances — useful for real-world SER research.

BADEM (2022)

Acted · Discrete emotions + neutral · 12 actors · 1,008 recordings · Audio-visual

Audio-visual acted emotion dataset with 1,008 recordings (84 videos per actor) from 12 actors (6M/6F). Each actor performed 12 sentences per discrete emotion plus neutral. Includes video for multimodal research.

  • Authors: Luisa Medina Fermino Carlos
  • Download: OSF Storage
  • Thesis: PDF

Automatic Speaker Verification (ASV)

Corpus CEFALA-1 (2019)

104 speakers · Audiovisual · Studio quality · 5 microphones

An audiovisual database of speakers designed for biometric, phonetic, and phonology studies. It contains recordings from 104 speakers (49 female and 55 male), and features a protocol guiding participants through a short monologue, paragraph reading, and sentence reading. Recorded in a professional studio using five different microphones—including a smartphone, a video camera, a wireless microphone, a lapel microphone, and a condenser microphone—the dataset is well-suited for speaker verification and biometric research tasks.

  • Authors: Arlindo Follador Neto, Adelino Pinheiro Silva, Hani Camille Yehia
  • Download: CEFALA
  • Paper: Rev. Estud. Ling.

BRSpeech-DF-Dataset

Real & synthetic speech · 459K samples · Brazilian & European Portuguese · Deepfake detection & speaker verification

459,000 audio samples of real and synthetic speech in Portuguese, covering both Brazilian and European variants. Synthetic samples are generated using multiple zero-shot TTS models, making the dataset suitable for deepfake speech detection and speaker verification research.

Gesture Synthesis

BRG-Unicamp

⚠️ Not yet publicly available

Brazilian Portuguese gesture synthesis dataset from AI-Unicamp.

Facial Synthesis

Unicamp Facial Synthesis Dataset

⚠️ Not yet publicly available

Facial synthesis dataset from Unicamp repository.

Lucas H. Ueda

Lucas H. Ueda

I’m a PosDoc Researcher with focus on speech synthesis and recognition.

Recent Posts

Brazilian Portuguese Speech Datasets
Visitor Counter
Visitor Counter