Start with the constraints
Missing Fundamental Labs starts with a practical question: what can a smaller audio model do well?
A limited parameter budget makes tradeoffs explicit. We want to study those tradeoffs carefully: what to preserve, what to simplify, and which changes actually improve the voice a listener hears.
Make the result inspectable
A score is much more useful when you can see how it was produced. We want our reports to describe the evaluation data, inference settings, hardware, and remaining weaknesses.
Listening examples should include difficult cases. Training descriptions should include the choices that did not work. Reproduction instructions should be specific enough to expose missing assumptions. These are the standards we want to work toward.
Build on open work
Kyutai’s open-science approach is an important inspiration for the lab. We value research that comes with the tools to use it, examine it, and extend it.
Missing Fundamental Labs is an independent project. We will identify the work we build on and keep upstream attribution and licensing visible.
Begin with Rivet
Our first focus is Rivet, our core text-to-speech model. This site will collect its project updates, technical findings, papers, and release documentation as they become available.