←Notes SMOREB
Leer en español NOTE

Ollama joins the voice models

Stiven Stiven Morelos·Oct 28, 2026 ·4 min read

This week Ollama started talking. Since version 0.40.11 you can turn text into speech on your own computer, without sending anything to a server. And the model behind it, Phonon, weighs 64 MB.

I’ve already listened to the voices and they sound really good. So this note is about what it is, what you can do with it today, and what I’m going to use it for.

What Phonon is

Phonon is a family of text-to-speech models from Gradium, released as open models on October 20. Ollama added it as its first speech model on the 24th.

What makes it special is the size. The English model has 31 million parameters and its file is about 64 MB. For comparison, a small language model weighs several gigabytes. On a Mac with an M5 Max, Ollama reports it uses around 121 MB of memory. It doesn’t need a dedicated graphics card: Gradium built it to run on a CPU, and in Ollama it uses the Mac’s chip. And once downloaded it works offline: neither the text nor the audio leaves your machine.

It speaks five languages: English, Spanish, French, German and Portuguese. There’s one model per language and a multilingual one with all five.

Watch the license

This is the part that matters most if you plan to use it in something you charge for:

ModelLanguagesLicenseCommercial use?
phonon:31m-enEnglishApache 2.0Yes
phonon:40m-enEnglishCC BY-NC 4.0No
phonon:40m-es (and fr, de, pt)One per languageCC BY-NC 4.0No
phonon:40m-multilingualAll fiveCC BY-NC 4.0No

In short: only the 31M English model is free for commercial use. The rest, Spanish included, are non-commercial only; to use them in a product you need to talk to Gradium. For personal projects you can use any of them.

Gradium also asks for something I think is healthy: don’t use it to impersonate real people, and don’t clone voices without consent.

How to try it

For now, in Ollama it only works on Macs with Apple Silicon. Windows and Linux are coming “soon”, with no date.

ollama pull phonon
ollama run phonon "Hello from Ollama."

# to list the available voices
ollama show phonon

And if you want to use it from your code, Ollama exposes an OpenAI-compatible endpoint:

curl http://localhost:11434/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "phonon", "input": "Your balance is up to date.", "voice": "freya"}' \
  --output voice.wav

If you already use OpenAI’s speech API, you change the URL and you’re done.

Outside Ollama, Gradium has its own ways to run it: on iPhone and Mac with Core ML, on Android, in the browser, in Python, with Docker, and on a Raspberry Pi 5, where they say it runs faster than real time.

How I’m going to use it

In my bots. I’m building my own bot system for my personal projects, and a local voice changes everything: I can talk to them without paying per sentence and without what I say leaving my home. Since it’s just for me, I can use the Spanish model.

In inventions. I want to give a voice to little things: small AMOLED screens with an ESP32-S3 and a Raspberry Pi 4. One technical detail here: the ESP32-S3 doesn’t have the memory to run the model. So the Raspberry (or a home server) generates the voice, and the ESP32 is the mouth and the face: it plays the audio and shows the animation on the screen. Gradium only confirms real time on the Pi 5, so I’ll have to test the Pi 4. When I do, I’ll write about it here.

In Virso, later on. The idea is an open integration, so anyone who wants to can give the assistant a voice with a local model. But the license rules: for anything commercial I can only use the English model, and most people who will use Virso speak Spanish. So I’m still thinking that one through.

Why I’m excited

Not long ago, giving a voice to a personal project almost always meant paying for an API or settling for a robotic voice. Today it’s 64 MB running on the computer you already own. That opens the door to a lot of experiments I wouldn’t even have tried before.

What would you make talk?

Sources: Ollama’s announcement ↗, Phonon at Gradium ↗, the model on Hugging Face ↗.