The House Learns to Talk Without Phoning Home

Ollama now does text to speech locally on Apple Silicon behind an OpenAI-compatible speech endpoint. For a house agent, the voice no longer has to leave the house. What shipped, what the licenses say, and what's next.

In this series I’ve been counting the computers in the house and asking what they do all day. The house is full of computers now. Last time, I wrote about running the house model on the quiet chip, so the box that runs all day draws less power and makes less noise.

There was still one thing that left the house every time: the voice. The model could think locally, but if you wanted it to answer out loud, the text usually went to a cloud speech API and came back as audio.

That just got easier to keep at home.

What Ollama shipped

Ollama published Speech models in Ollama on October 24. Ollama can now turn text into speech. The first model is Phonon, a family of small text-to-speech models from Gradium.

The details that matter for a house box:

  • Platform: Apple Silicon Macs, on Ollama 0.40.11 or later. Windows and Linux are listed as coming soon.
  • Same endpoint shape: Phonon works with the OpenAI-compatible /v1/audio/speech endpoint. Per Ollama, existing code that uses the OpenAI SDK only needs a new base URL.
  • Streaming: the response streams while the audio is generated, so playback can start right away. Pipe a chat model’s reply into Phonon and it starts speaking after the first sentence.
  • Format: 24 kHz mono, WAV by default or raw 16-bit PCM.
  • Voices: picked by name from Gradium’s catalog, and ollama show phonon lists them.

The swap is about this small:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")

The numbers, and their conditions

Ollama reports that on a MacBook Pro with M5 Max, it was 35% faster than the next fastest engine, Gradium’s own runtime. The conditions are in the fine print: generating 45 seconds of speech, median of 10 runs, with Gradium’s runtime using its own mix of precisions and Ollama running the model in bf16.

On accuracy, Ollama says its implementation matches the word error rate of the reference implementation in Hugging Face Transformers, on 200 English sentences from the Seed-TTS test set, transcribed with Whisper large-v3.

That’s one machine, and it’s a fast one. The box in your house may be something older and slower. I’m not going to claim a latency number for my house until I’ve measured one.

Read the license tag

This is the part I’d check before wiring it in. Per Ollama:

  • phonon, the English model, is Apache 2.0.
  • The 40M models for French, German, Spanish, Portuguese and the multilingual one are CC BY-NC 4.0, which limits them to noncommercial use.

For a house agent that only ever talks to your family, that may be fine. For anything you sell, or anything a business runs, pull the exact tag and read its license first. Same model family, different terms.

In my sketch of the setup, the speaker sits on a stack of three old phone books.

What’s next: talking and listening at once

Phonon covers the talking half. The house still has to hear you, and today that means a separate speech-to-text step before the model ever sees your words.

Block’s Dessa points at where this goes. Block describes it as a speech-to-speech model trained only on human conversation, full-duplex so it listens and talks at the same time, and says it runs locally on a MacBook Pro M4 Max or better. You can’t use it yet. Per TechCrunch, there’s no public access, and Block’s page says it’s giving early access to a small group of AI safety researchers.

So for now, the practical house setup is still a pipeline: hear, think, speak. The difference is that the last step no longer has to phone home.

How I’d test it in the house

Same approach as the quiet chip. Pick the phrases the house actually says, like “garage door is open” or a weather summary, and run them on the box you’d really use.

phonon test, house box, 20 phrases
  time to first audio:  __ s
  full reply length:    __ s
  power under load:     __ W
  understood on first listen: __ / 20

If the box can do that while the model keeps answering, the house can talk without anything leaving the network. If it can’t, you’ve learned that from a notebook, not from a family complaint.

This also fits the rule I’ve been using for the rest of the house: know what every box does, and know what it sends out. A voice that stays local is one less thing to answer for.

Sources

Frequently asked questions

Can Ollama do text to speech?

Yes. As of Ollama 0.40.11, announced October 24, 2026, Ollama runs speech models, starting with Phonon from Gradium. It runs on Apple Silicon Macs for now, with Windows and Linux support listed as coming soon.

Does Ollama's text to speech work with the OpenAI SDK?

Yes. Phonon works with the OpenAI-compatible /v1/audio/speech endpoint, so existing OpenAI SDK code only needs a new base URL pointing at the local Ollama server. Audio is 24 kHz mono, returned as WAV by default or raw 16-bit PCM, and it streams while it is generated.

Can I use Ollama's Phonon voices commercially?

The English phonon model is released under Apache 2.0. The 40M language models for French, German, Spanish, Portuguese and multilingual are released under CC BY-NC 4.0, which limits them to noncommercial use. Check the license of the exact model tag you pull.

Disagree or have a war story? Reply on X