Ollama adds local speech generation on Apple Silicon Macs with Gradium's Phonon models

Ollama can now generate speech, its first move beyond text and images. The feature launched on 24 October with Phonon, a family of small text-to-speech models from Gradium, and works on Macs for now, with Windows and Linux support described as coming soon. It needs Apple Silicon and Ollama 0.40.11 or later, although the model page lists 0.40.9 as the minimum.

Audio is produced in real time, so a chat model's reply can be heard while it is still being written. Users run ollama run phonon with some text, choose accents such as American or British English with a voice flag, and either pipe output from a chat model or save a WAV file. An OpenAI-compatible speech endpoint means code written for OpenAI's SDK only needs a new base URL.

Ollama says its implementation, built on Apple's MLX, ran 35% faster than Gradium's own runtime on a MacBook Pro with an M5 Max, while matching the word error rate of the reference setup in Hugging Face Transformers. There are English, French, German, Spanish and Portuguese models, each with its own variant plus one multilingual model.

Licences differ by checkpoint. The default 31M English model is under Apache 2.0, while the 40M language models are CC BY-NC 4.0 and limited to non-commercial use. One analyst warns the default may offer only a single voice. Ollama has nearly 185,000 GitHub stars.