Ollama adds Phonon speech on Macs, with different licences for model variants

Ollama’s October 24 release adds text‑to‑speech support through Phonon, a model family from Gradium. The announced speech runtime is available on macOS with Apple Silicon; Windows and Linux support are described as coming soon. This is text turned into audio, a different task from transcribing a recording.

The announcement recommends Ollama 0.40.11 or later and shows ollama pull phonon followed by ollama run phonon with text to speak. The model page lists an earlier minimum, 0.40.9. Using the announcement’s newer recommendation avoids presenting the older number as an independently tested minimum.

The current model page identifies the default phonon model as Phonon 31M English, with one voice named default. It lists explicit 40M variants for English, French, German, Spanish and Portuguese, plus a multilingual variant covering those five languages.

The release’s general voice examples use named voices from Gradium’s catalog, but that does not establish that those voices work with the default 31M checkpoint. For the default package, the model page’s single‑voice description is the narrower supported statement. Selecting a language variant and selecting a voice should not be treated as interchangeable choices.

Both the release and model page distinguish the licences: the 31M English checkpoint is identified as Apache 2.0, while the 40M English, French, German, Spanish, Portuguese and multilingual checkpoints are identified as CC BY‑NC 4.0, limited to noncommercial use.

That distinction travels with the selected checkpoint. A multilingual or larger English variant should not be described as having the same published licence as the default simply because both are served through Ollama. These are the publisher’s licence labels; this article does not provide a broader rights clearance for generated voices or a particular deployment.

The release documents an OpenAI‑compatible /v1/audio/speech endpoint at the local Ollama server. Its examples provide the model and text input, with a voice parameter where applicable. The response can stream as the audio is generated.

The documented output is 24 kHz mono, returned as WAV by default. Setting response_format to pcm requests raw 16‑bit samples instead. The model page also advises supplying text intended to be spoken: formatting such as Markdown headings may be read aloud. The useful setup decision is therefore the supported Mac runtime, exact checkpoint and licence, then the voice and audio format needed by the application.

AI‑generated conceptual illustration by OpenTools Team: written text becomes audio beside an unbranded laptop. This is an editorial metaphor, not Ollama’s interface, a real product photograph or evidence of a tested model.