Ollama now speaks, locally. Only its English model allows commercial use, so read the license first

The most important line in Ollama’s text-to-speech announcement is not the benchmark. It is this one: “The 40M models are released under CC BY-NC 4.0, which limits them to noncommercial use.”

On October 24, Ollama added speech generation, starting with the Phonon family from Gradium. The default English model, phonon, is licensed under Apache 2.0. The French, German, Spanish and Portuguese variants, and the multilingual one that covers all five languages, are noncommercial. A company that builds a customer-facing voice feature in Spanish or Portuguese on these models would be using them outside their license.

That does not make the release less interesting. It makes the order of evaluation clear.

What shipped

• Local speech generation with Ollama 0.40.11 or later, on Apple Silicon Macs only for now, using Apple’s MLX framework. Windows and Linux are “coming soon.”
• Streaming: audio plays while it is generated, so a piped chat reply starts speaking after its first sentence.
• An OpenAI-compatible endpoint, /v1/audio/speech on the local server. Existing code written for OpenAI’s SDK “only needs a new base URL.”
• Voices from Gradium’s catalog, selected per request, and 24 kHz mono output as WAV or raw PCM.
• A vendor benchmark: on a MacBook Pro with an M5 Max, Ollama says it was 35% faster than the next fastest engine it tested, Gradium’s own runtime, generating 45 seconds of speech (median of ten runs), with word error rate matching the reference implementation on 200 English sentences. One machine, one comparison, measured by Ollama.

Why local speech matters to enterprises

Voice features have usually meant a cloud speech service next to whatever model handles the conversation. That adds a processor for every utterance, a per-character bill and a network round trip in the middle of a real-time exchange. We described voice agents as real-time systems, where latency shapes the architecture more than intelligence does. Local synthesis removes one hop and one data flow.

The OpenAI-compatible endpoint is the practical part. A team can prototype against a cloud speech API and switch to local generation by changing a base URL, or keep both behind the same client and route by use case. That is the same portability argument as for text models, applied to audio.

Four questions, in this order

1. Is the use commercial? If the answer is yes and the language is not English, the current Phonon variants are out. Internal tools are a gray area that your legal team, not your engineers, should decide.
2. Where will it run? Today, only on Apple Silicon. That suits developer machines and some edge kiosks; it does not suit the Linux servers most companies use for production inference.
3. Is text-to-speech the right shape? Phonon turns text into speech. It does not listen, and it does not handle interruptions or turn-taking. A conversational voice agent still needs speech recognition and a model in the loop.
4. Does local mean compliant? Local processing keeps audio off a third party’s servers, which helps. It does not by itself satisfy retention, consent or security requirements, and a synthetic voice still raises the question we discussed for AI avatars: whose voice may the company synthesize, and does the listener know it is synthetic?

Where it fits today

The realistic enterprise uses now are internal and English-first: accessibility features in internal tools, spoken summaries for field teams on Macs, prototypes of voice interfaces that will later move to a licensed production model. For workloads that belong on a machine under the desk, local speech completes the stack.

For multilingual customer-facing voice, the release is a signal rather than a solution: speech is becoming a standard part of the local model runtime, behind the same API shape as the cloud. When Windows and Linux support arrive, and if commercially licensed models in other languages follow, the case gets much stronger. Until then, the license decides before the latency does.