Research.
Speech recognition, programming languages and the systems that turn computation into useful software.
Speech &
machine learning
The research behind avette starts with a practical problem: making speech recognition work well with the way people actually talk, in more languages, on their own devices.
Data that sounds like people
We are developing a speech corpus with consent, with collection in Europe. The goal is training data that reflects real speech, rather than relying only on audiobooks or parliamentary recordings.
From training to the device
Our work includes fine-tuning open speech models, GPU training with CUDA, and running recognition locally. Current applications are dictation on the Mac; further work includes iPhone input, transcription and live translation.
Research in progress. A paper on the speech recognition work is in draft.
Languages &
computation
vix explores how to describe computation so that dependencies, reuse and execution are part of the system, rather than conventions every project has to reinvent.
One changed input, only the dependent work repeated. A conceptual illustration, not a benchmark.
Beyond a build system
A language and platform for general incremental recomputation. Build systems and faster CI are applications of that idea, not its boundary.
Correctness across machines
The work spans dependency tracking, caching, sandboxed execution and distributed builds. Rust and memory safety are a foundation; we also want correctness in the way work is described and reused.
More about vix