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 422555786 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.

SemFence PT-BR Synthetic Speech

24.7 horas de fala sintética em português do Brasil, com duas vozes (uma feminina e uma masculina) lendo as mesmas 10,675 frases do banco de frases aberto da SemFence. Feito para treinar modelos de voz pequenos e abertos, como o tiny-voca, em português. 24.7 hours of synthetic Brazilian Portuguese speech, with two voices (one female, one male) reading the same 10,675 sentences of the SemFence PT-BR sentence bank. Built to train small open TTS models in Portuguese.

Created by SemFence · researcher: Gustavo Toledo.

Clips 20,442 (10,207 female, 10,235 male)
Duration 24.7 hours (median clip 4.3 s, max 10.8 s)
Voices F_s4 (female), M_sm4 (male), both synthetic, Brazilian accent
Audio 24 kHz mono, 16-bit FLAC, stored inside the Parquet files
Language Portuguese (Brazil)
License CC BY 4.0, credit required (see "License and attribution")

Usage

from datasets import load_dataset

ds = load_dataset("semfence/ptbr-synthetic-speech")
row = ds["train"][0]
print(row["voice_id"], row["text"])
audio = row["audio"]["array"]      # numpy array, 24 kHz

Splits

Splits are made by sentence: both voices reading a sentence are always in the same split.

Split Clips Hours
train 19,025 23.0
validation 818 1.0
test 599 0.7

How it was made

  1. The two voices were created from text descriptions with Qwen3-TTS 1.7B VoiceDesign. They are not recordings or imitations of real people. The reference clips are in references/, and the descriptions and seed are in voices.json.
  2. Each sentence was read by both voices with Chatterbox Multilingual, Brazilian Portuguese (Resemble AI, MIT), using each reference clip as the voice prompt (exaggeration 0.5, temperature 0.8, cfg_weight 0.5).
  3. Each clip was transcribed with Whisper (openai/whisper-large-v3-turbo, Portuguese). A clip was generated again (up to two attempts, with a different seed) if the word error rate was above 15%, the speaking rate was outside 1–5 words per second, or the clip was longer than 25 s. Clips that failed twice were left out: 468 female and 440 male (about 4.3%). They are listed in generation/review.jsonl (the audio was not kept).
  4. Accepted clips were peak-normalized to 0.95.

The word error rate compares the Whisper transcript with the sentence after numbers are expanded into words and accents, case and punctuation are ignored. Mean WER of the accepted clips: 0.66%; 296 clips (1.4%) needed a second attempt.

The scripts that built the sentences and the audio, including the LLM prompts, are in generation/scripts/.

Listening check. The automatic screening catches wrong or missing words. It does not measure naturalness or accent. The authors listened to a random sample of 10 clips (5 sentences, both voices) and found the accent, the pronunciation of numbers and times, and the audio quality correct and clean. There was no formal listening test.

Columns

Column Description
id Clip ID, <text_id>_<voice_id>
text_id, voice_id, voice_gender Sentence ID (same as in the sentence bank), voice (F_s4 or M_sm4), and female or male
text The sentence that was read
split, category, origin Split, kind of sentence, and where the text came from
source_license, source_url License of the text at its source, and a link
audio 24 kHz audio
duration_seconds, sample_rate Audio length and rate
asr_transcript, normalized_wer Whisper transcript and its word error rate against text
attempt, seed Generation attempt (1 or 2) and the random seed used
teacher_model, asr_model, reference_audio Model that read the text, model that checked it, and the voice prompt

Text origins: llm 10,685, common_voice 7,813, template 1,944 clips.

Limitations

  • The speech is synthetic. It carries the pronunciation habits and errors of the model that generated it. The data cannot be more natural than Chatterbox itself.
  • Two voices, one accent. Brazilian Portuguese; the voice descriptions asked for a Brazilian accent and no specific region. Read speech only, no spontaneous speech or background noise. Many sentences are short.
  • Watermark. Chatterbox embeds an inaudible Perth watermark in the audio it generates, and it is present in these clips.
  • Text. The llm sentences were not reviewed by a person, and the filter for European Portuguese in the Common Voice sentences is automatic. See the sentence bank card for details.
  • Screening is automatic. A short sentence can pass with one wrong word, since 15% of a 6-word sentence is less than one word. Use normalized_wer to filter more strictly.

License and attribution

This dataset is released under CC BY 4.0. You may use, share and adapt it, including commercially, as long as you give credit:

SemFence PT-BR Synthetic Speech, by SemFence (Gustavo Toledo) — https://huggingface.co/datasets/semfence/ptbr-synthetic-speech

Please keep this credit in your dataset card, model card, paper or repository when you use the data or train a model on it. The sentences from Common Voice are CC0 at their source; the compilation, the audio and everything else here are CC BY 4.0.

Citation

@misc{semfence2026ptbrsyntheticspeech,
  author       = {{SemFence} and Toledo, Gustavo},
  title        = {SemFence PT-BR Synthetic Speech: two synthetic Brazilian Portuguese voices},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/semfence/ptbr-synthetic-speech}
}

Acknowledgements

  • Chatterbox Multilingual pt-BR by Resemble AI (MIT) read all the sentences.
  • Qwen3-TTS VoiceDesign and Qwen3-8B by Alibaba Qwen (Apache-2.0) designed the voices and wrote the conversational sentences.
  • Whisper large-v3-turbo by OpenAI (MIT) checked every clip.
  • Mozilla Common Voice (sentence collector, CC0) provided part of the text.
  • The recipe (synthetic voices designed from text, screening by ASR, open data) follows tiny-voca by Muhammed Nazeem.
Downloads last month
101

Models trained or fine-tuned on semfence/ptbr-synthetic-speech