SOTAQUE: Our Open Source Dataset of Brazilian Voices
Published on
I want to share a new project I just put online: SOTAQUE (“accent”, in Portuguese).
Here’s the site: sotaque.ia.br
The idea is pretty simple, but I think it can have a really nice impact: creating an open dataset of Brazilian Portuguese voices.

If you’ve ever used AI tools for transcription, voice assistants or speech generation in Portuguese, you’ve probably noticed two things.
The first is that a lot of AIs speak a kind of “pasteurized” Portuguese, as if everybody sounded the same.
The second is that, when a stronger accent comes in, performance often gets worse. The transcription makes more mistakes, comprehension drops, and the experience just isn’t as good.
In my view, these two problems have a similar root: there’s a lack of open, diverse data of Brazilian voices.
And that’s where SOTAQUE was born.
The project’s proposal is to put together a public collection of recordings from people of different regions, ages and backgrounds, to help develop voice technologies that are more representative of real Brazilian Portuguese.
When I say voice technologies, I’m talking about things like:
- transcription models
- voice assistants
- speech synthesizers
- audiobooks
- educational tools
- research and open-source projects
One of the things that bothers me the most in this space is that the data infrastructure is often closed, concentrated in a few companies, and without much transparency. That makes life pretty hard for anyone who wants to research, test or build better solutions in Portuguese.
And in Brazil’s case it weighs even more, because our diversity of speech is huge.
There’s no such thing as “the Brazilian accent”.
There’s the São Paulo accent, the Rio one, the Minas one, the Bahian one, northern, gaúcho, caipira, and many others. And that’s without even getting into age, social context, speech rhythm, vocabulary and local expressions.
If the training data doesn’t reflect that diversity, the technology won’t reflect it either.

SOTAQUE was born precisely to try to help with that.
Contributing is meant to be simple: you can record or upload an audio clip, fill in some basic information and authorize its use under the consent form. From there, the goal is to build this dataset step by step, with curation, in the open, so it can be truly useful.
I think this kind of project is important because it helps not only to improve models, but also to strengthen the local ecosystem.
Researchers, Brazilian startups, teachers, creators and independent developers could benefit a lot more if better open datasets in Portuguese existed.
Anyway, I wanted to put the project online because I think this is a piece of infrastructure that’s still missing.
If you want to get to know it better or contribute your voice, the link is right here:
If it works out, it could become a really interesting dataset for improving how AI listens to and speaks Brazilian Portuguese.