Why we're building small speech models for cloud, servers and devices

We're a small team in Madrid building Lokutor. We want voice agents to work where people need them: in the cloud, inside a company's own servers, and on devices that cannot depend on a connection.

That means paying attention to the speech models themselves, not only the orchestration around them. Size, deployment and what happens during a conversation all matter.

Our platform work covers four parts of that conversation:

• Psst cleans the caller's audio before speech recognition. We build it for the recogniser behind it, not just for audio that sounds cleaner.

• Turno handles turn taking. It helps tell a real interruption from a short acknowledgement or the agent's own echo.

• Conv turns speech into text. Production currently uses Conv 1.0, based on NVIDIA's open Parakeet model. Our own Conv 2.0 is the next release.

• Versa 2.0 turns text into speech and supports voice cloning. It is already serving calls on our platform.

The language model is separate. We use an external provider for that part of the conversation; we are not claiming to have trained the whole stack ourselves.

CPU capability and our current cloud setup are different things. Psst and Turno run on CPU. Versa also runs on CPU, while our cloud currently uses an NVIDIA T4 for synthesis and recognition. We choose the hardware for the deployment rather than claiming every production call is GPU free.

At the smaller end, we're also building Ito for speech synthesis and Oído for speech recognition on the ESP32 S3. These are separate edge projects, not replacements for Conv and Versa in the cloud pipeline. Oído combines an NVIDIA acoustic model with our own engine, language model and chip port. Our current Ito release is English only. Its firmware audio has been checked against the host engine in emulation; physical board performance is still something we need to measure.

That's the work we're sharing: building the models we own, being clear about the open models we adapt, and checking what actually works on each target.

You can try the voice demo at lokutor.com without signing up. We'd like to hear which deployment constraint matters most in what you're building, and we're happy to go deeper on the models or the engineering.