Ollama’s new speech support makes a local assistant audible, but the model name alone no longer tells developers which capabilities or permissions they are getting. The October 24 release introduces Gradium’s Phonon family on Apple Silicon Macs. Its default English checkpoint has an Apache 2.0 license; the larger language-specific and multilingual checkpoints use a noncommercial license.
That split matters for anyone moving from a personal spoken assistant to a product. Selecting another language changes more than pronunciation. It can change the terms under which the model may be used, even though the same Ollama commands and speech endpoint provide the interface.
Phonon generates audio incrementally. Ollama can speak incoming text before the complete response has arrived, so a text model’s answer can feed directly into speech. This adds an output stage to a local assistant rather than replacing its reasoning model. Phonon is a text-to-speech model, not a model that listens to the user or decides what to say.
One family, different checkpoints
Ollama’s model page identifies the default phonon package with Gradium’s 31M English checkpoint. Gradium’s model card also specifies that its parameter count excludes the audio tokenizer. That makes the number useful for identifying the model, but insufficient as a complete description of the software or memory involved in producing speech.
The release offers French, German, Spanish and Portuguese through 40M checkpoints, alongside a 40M English model and a multilingual version covering all five languages. The important comparison is compact enough to keep beside the model choice:
| Checkpoint | Language coverage | Published license |
|---|---|---|
| Phonon 31M English, Ollama’s default | English | Apache 2.0 |
| Phonon 40M language-specific models | English, French, German, Spanish or Portuguese | CC BY-NC 4.0 |
| Phonon 40M multilingual | All five languages | CC BY-NC 4.0 |
The runtime’s open-source license is separate from those checkpoint licenses. Gradium’s repository explicitly says that model weights and voices have their own terms. A permissively licensed application does not turn every model or voice it loads into a permissively licensed asset. For a commercial project, the practical consequence is to record the exact checkpoint and voice, rather than approve the family name once and assume future substitutions are equivalent.
Speech becomes another local output
Ollama’s terminal integration supports both playback and saved files. A normal speech command plays audio; redirecting its output writes a WAV file. Piping a text model’s output into Phonon lets playback begin as text arrives. Those are useful distinctions for an assistant that sometimes reads a reply aloud and sometimes produces narration for later use.
Applications can instead use the OpenAI-compatible /v1/audio/speech endpoint. Ollama documents streamed WAV output and raw PCM, both at 24 kHz in mono. Its speech guide sets a 4,096-character input limit per API request, while the command-line path splits longer text at sentence and paragraph boundaries. A program handling a long document therefore needs to account for which interface it uses; terminal convenience is not automatically an API feature.
Compatibility also has a narrower meaning than accepting every speech option. The guide says unsupported formats, speeds or instructions return an error. Chat and embedding requests sent to a speech model also fail. An existing integration can reuse a familiar speech API shape without assuming that every option from another provider transfers unchanged.
The release announcement limits this Ollama rollout to macOS on Apple Silicon, with Windows and Linux support still forthcoming. Gradium’s own runtime documents broader platform integrations, which is a different availability claim. Readers following SekaiTech’s earlier coverage of local Ubuntu dictation should also distinguish the direction of the task: dictation turns incoming speech into text; this release turns outgoing text into audio.
Ollama recommends version 0.40.11 or later in its announcement, although its speech guide and model page list 0.40.9. The newer announcement provides the more conservative starting point. SekaiTech compared the release, current documentation and Gradium’s model information; it did not run listening tests or benchmark generation speed. For this release, a useful model record includes the checkpoint’s language and license, the selected voice and the output interface. Those details determine what the new speech stage can actually contribute to an application.