Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 665371995 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

tiny-broca dataset

A synthetic Brazilian Portuguese speech corpus with two voices, F1 (female) and M1 (male), made to train the tiny-broca text-to-speech model. Every sentence is read by both voices.

Sentences 9,384
Clips 16,434 accepted of 19,504 generated
Duration 17.3 hours
Voices F1, M1
Audio 24 kHz mono, 16-bit FLAC, stored inside the Parquet files
Language Portuguese (Brazil)
License CC BY 4.0

Usage

from datasets import load_dataset

batch1 = load_dataset("brenorb/tiny-broca", "batch1")
batch2 = load_dataset("brenorb/tiny-broca", "batch2")
batch3 = load_dataset("brenorb/tiny-broca", "batch3")

row = batch1["train"][0]
print(row["voice"], row["text"])
audio = row["audio"]["array"]

Contents

Counts are the accepted clips' unique sentences; each sentence is read by both voices, so the clip count is roughly twice the sentence count.

Config Split Sentences Clips 16,434 accepted of 19,504 generated
batch1 train 3,560 7,119 conversational sentences and Common Voice
batch1 validation 647 1,294 same
batch1 test 404 807 same
batch2 train 3,327 6,653 Common Voice and templates
batch3 train 343 561 money phrases: amounts, cards and bills

batch1 mirrors the first batch of the corpus this project follows (conversational speech plus natural sentences); batch2 is training data only, mirroring its second batch (news-style sentences plus sentences that exercise every rule of the text normalisation). batch3 is training data only: phrases dominated by money amounts, generated after the first training run so the model hears numbers read at their full value. Splits are made by sentence: both readings of a sentence are always in the same split, so the validation and test sentences are the same for every model trained on this corpus.

How it was made

  1. The two voices were created by voice cloning: each is Qwen3-TTS 1.7B prompted with a short clean recording of a real speaker from Mozilla Common Voice (CC0). Voices designed from text descriptions (the English model's method) were auditioned and rejected as less natural. The reference clips are in references/, with their sources and seeds in voices.json.
  2. Each sentence was synthesised in both voices with Qwen3-TTS 1.7B Base, using the reference clips as the voice prompt.
  3. Each clip was transcribed with Whisper (multilingual, pt). Clips above 15% word error rate, or with implausible duration, speaking rate or silence, were generated a second time; clips that failed twice were left out. 0 clips failed twice and are listed in generation/*/review.jsonl.
  4. Accepted clips were peak-normalised to 0.95.

The generation scripts and their settings are in generation/.

The automatic screening catches wrong or missing words. It does not measure naturalness, and the corpus has not been checked by listeners clip by clip.

Text sources

The source column records where every sentence came from.

Source Sentences License of the text
Mozilla Common Voice (Brazilian Portuguese) 4,405 CC0
Conversational sentences, generated deterministically from written slot pools 3,500 this project
Templates: Brazilian numbers, dates, times, money, identifiers, questions 1,079 this project
Money phrases (batch3) 400 this project

Accepted clips by source: Common Voice 7,706, conversational 6,589, templates 1,578, money 561.

The conversational slots keep the corpus reproducible: the same seed produces the same sentences.

Columns

Column Description
id Clip ID: batch, voice, split and sentence index, e.g. b1_F1_test_00390
text_id Sentence ID in the corpus, e.g. broca01_00000
batch 1 (conversational and Common Voice) or 2 (Common Voice and templates)
split train, validation or test
text The sentence that was read
source cv, converse or templates
source_license, source_url cc0-1.0 for Common Voice sentences, otherwise project-own
voice F1 or M1
audio 24 kHz audio, as bytes, path and sample rate
duration_seconds, sample_rate Audio length and rate
asr_transcript, normalized_wer Whisper transcript and its word error rate against text
gen_seconds, rtf Generation time and real-time factor of the clip
teacher_model, reference_audio Model and voice prompt used

Annotations

annotations/ holds the letter-level annotations used for training: the letter tokens, the stress feature, and the duration of each letter in 10.7 ms frames, from forced alignment with jonatasgrosman/wav2vec2-xls-r-1b-portuguese. Clips that fail the alignment checks are listed with a reason in the *.rejected.jsonl files. See docs/annotations.md in the code repository. 14,290 clips are annotated; 933 failed the alignment checks.

Limitations

  • The speech is synthetic. It carries the pronunciation habits and errors of the model that generated it (for example, a reduced vowel in the standalone word se in some clips).
  • Two voices, one accent, read speech only.
  • The two voice references are real recordings of Common Voice contributors; they are included as references/ for provenance and are not training audio.

License

CC BY 4.0. Please credit "tiny-broca dataset, brenorb" and link to this page.

@misc{brenorb2026tinybrocadata,
  author = {brenorb},
  title  = {tiny-broca dataset: synthetic two-voice Brazilian Portuguese speech},
  year   = {2026},
  url    = {https://huggingface.co/datasets/brenorb/tiny-broca}
}

The Common Voice sentences are CC0, from https://github.com/common-voice/common-voice. The audio was generated with Qwen3-TTS (Apache-2.0). See LICENSE.md.

Downloads last month
1

Models trained or fine-tuned on brenorb/tiny-broca