A voice of its own.
Rivet is the text-to-speech project at the heart of Missing Fundamental Labs. We’re exploring how much expression and fidelity we can retain in a smaller model, and what makes that model useful outside a research environment.
This page introduces our research direction. Technical specifications, listening examples, and measured results will be added with the release.
The questions we’re asking
Model size is one part of the problem. We want to understand the relationship between capacity, data, and the details that make a voice feel natural.
- How does reducing model capacity affect intelligibility, rhythm, and expression?
- Which training and data choices matter most under a constrained compute budget?
- How should we measure inference speed and memory use on the hardware people actually have?
- What evidence does another researcher need to reproduce the result?
What a release should include
Our aim is to publish the evidence alongside the model: a technical report, documented evaluation, clearly identified audio samples, and the resources needed to understand the work.
Availability of code, weights, and any redistributable data will be stated explicitly, with a license attached to each artifact.