Docs Book a Demo Sign in
Models

We train our own speech models.

Versa, our speech foundation model, gives every agent its voice, in thirteen languages. Turno decides whose turn it is, Psst hears the voice and not the room, and Conv 2.0 is our recogniser. All of them are small models by industry standards.

The models

Five models, one stack.

  • Versa 2.0

    Speech synthesis Speaks on every call

    Our speech foundation model. It generates speech one 70 ms frame at a time, for every stream in flight at once. Thirteen languages: nine in ten voices, and Dutch, Polish, Romanian and Russian with a female and a male voice each. Any new voice from a few seconds of reference audio.

    First audio in about 25 ms for one request and a median of about 90 ms with 32 simultaneous streams on one GPU. Held-out sentences read back by a recogniser score 9.62% mean word error, against 10.42% for real recordings of the same sentences.

    120.0 million deployable parameters · Text to speech →

  • Turno

    Turn-taking and barge-in Decides every turn

    Decides on the audio itself whether the caller has finished, is interrupting, or is only saying “mhm”, instead of waiting on a silence timer.

    Average precision 0.931 against 0.876 for Silero VAD on a neutral real-world test set (TEN-VAD leads on voice activity alone at 0.985), flat from clean audio down to 0 dB, at about half Silero’s cost per inference.

    18,651 parameters · Turno benchmark →

  • Psst

    Noise suppression Cleans every call

    Removes background noise and other voices before recognition, tuned for what matters to a voice agent: the words, not how pretty the audio sounds.

    Word error after cleaning 36% lower than DeepFilterNet3, the strongest published real-time denoiser, and the only system we tested that improves on no denoising at all. One CPU core cleans 18 calls at once.

    2.56 million parameters, 32 ms latency · Noise cancellation →

  • Conv 2.0

    Speech recognition Listens on every call

    Our own recogniser: one model for nine languages, Catalan, Galician and Basque included, about a fifth smaller than the NVIDIA Parakeet model it replaced. It has transcribed every production call since 20 October 2026.

    6.05% word error on an English conversational benchmark scored like Artificial Analysis’s, 9.62% across its eight other languages, as served today (5.82% and 7.56% with its language model, not yet in production); about 55 ms to transcribe a typical utterance on our GPUs.

    About 466 million parameters · Conv benchmark →

  • Ito

    Speech synthesis for microcontrollers Research, open weights

    A streaming neural voice built for the ESP32-S3, a $5 chip with no neural accelerator. English, two voices, no cloud.

    Scores 4.44 on UTMOS22 with 0.4% word error on 54 held-out prompts. Verified bit for bit in Espressif’s emulator; no physical board has run it yet.

    4.4 million parameters, 4.9 MB · The Ito post →

Turno

Turno decides whose turn it is.

The turn passes on what it hears, not on a silence timer. A “mhm” doesn’t take it. A real interruption does.

Psst

Psst hears the voice, not the room.

The same two seconds of speech, drawn from the real audio. Drag the line to compare, or press play and watch it move.

Why small

Small models go further.

  • PriceLess compute per call is how a voice agent costs about 6¢ a minute, language model included.
  • AnywhereThe same models run in our cloud, on your servers or inside a device. Running on an ordinary CPU is the proof, not the point.
  • SpeedA small model answers fast: Versa’s first audio in about 65 ms, and a voice agent’s reply about 0.6 s after the caller stops talking.
What isn’t ours

The parts we don’t build.

  • Language modelFrom low-latency providers, plugged into our pipeline.
  • OídoOur inference engine for the ESP32-S3 runs NVIDIA’s open Conformer-CTC Small. The engine is ours; the network is NVIDIA’s, and our Spanish version is a fine-tune of it.
FAQ

Common questions.

Does Lokutor build its own AI models?

Yes. We design and train the speech models that do the listening and the speaking: Versa for the voice, Turno for turn-taking, Psst for noise suppression, Conv 2.0 for speech recognition and Ito for microcontrollers. The language model comes from a provider.

Why are Lokutor’s models so small?

Because a model that is cheap to run can be run anywhere and sold for less. Versa has 120.0 million deployable parameters and Turno 18,651, small by industry standards. Our noise suppression runs on an ordinary CPU in production, and we deploy the models on a customer’s own servers or on a device, scoped case by case.

Do Lokutor’s models need a GPU?

Our hosted API serves Versa and our recogniser on GPUs, because one GPU carries many simultaneous calls: Versa’s first audio takes about 25 ms for one request and about 90 ms with 32 streams at once. Nobody has measured the current Versa on a CPU, so we do not claim it runs on one. Our noise suppression does run on a CPU in production.

Where are the models trained?

On cloud GPUs and on the Barcelona Supercomputing Center’s cluster, through its AI Factory program for European AI startups.

Use them in your product.

Through our API, on your own servers, or inside your device.