Our own models, voice first
We started with speech because it is the surface with the least tolerance for latency and the least room to hide.
Building our own models was a decision about control: over latency, over cost, over what the model is allowed to do, and over where it runs.
Speech was the right place to start because it is unforgiving.
Why voice first
Real-time speech has no room for a slow pipeline and no opportunity to revise. Every weakness is immediately audible: latency, turn-taking, handling interruption, recovering from a mis-hearing.
A system that holds up there tends to hold up elsewhere.
What we built
A speech stack trained and served in-house: recognition tuned for the conditions our deployments actually encounter, and synthesis that holds a consistent voice without drifting across a long exchange.
Details are in the Ruby Voice system card.
Evaluating on real conditions
Clean-audio benchmarks flatter every system. We evaluate on noise, crosstalk, accents, and degraded connections, because those are the conditions the model meets in deployment.
Where it runs
On our own GPUs, under our own serving stack, which is what makes the local enterprise deployment possible at all.
What comes next
Voice was the first surface, not the goal. The work now extends into the reasoning models that do the rest of the job, covered in expanding into broader research models.
More in Enterprise