We have published Cadence, the workbench where we train and evaluate the models behind Oasis Speech. The public repository is a fork of our production system, under the MIT license. People and coding agents work in it through the same API.
Every month, for many languages, we fine-tune, measure and compare the models behind our speech recognition. Cadence is where that happens. Today the repository is public: github.com/usunrise88/cadence.
What it is
Cadence covers the whole life of a speech model: data, training, evaluation, export and deployment, and the loop that brings production calls back into training data.
- Agent-native. A coding agent works in Cadence through the same API as the interface. Its edits arrive as drafts under its own name. Anything that spends money or changes a shared asset waits for a person’s approval.
- API-first. One OpenAPI contract generates the server, the web client, the agent tools and the command line, so an action has one name everywhere.
- Gates. Cadence evaluates a model against frozen evaluation sets through the real streaming decoder, and registers it only after it passes its gate.
- Clean tests. Cadence checks every training input against every evaluation set, so an evaluation clip cannot leak into training.
- Budgets. Each project and each agent session has a GPU budget. Cadence counts queued work against it, so a queue of small jobs cannot add up past the limit.
A fork of what we run
The public repository is a fork of the system we use in production. It is the same workbench we trained Oasis Speech 2 in, published under the MIT license: use it, change it and build on it.
Read more
- Cadence: the workbench we train Oasis Speech in, the research note with more detail on how it works.
- Oasis Speech 2 is live: our agents now hear 48 languages.
- The repository: github.com/usunrise88/cadence.