Ollama for Mac: Local text-to-speech now available in German too

Ollama can talk now. The popular app for running AI language models locally supports text-to-speech for the first time as of version 0.40.11 – for now only on Macs with Apple Silicon. The Ollama team relies on Phonon, a family of small speech models from the French start-up Gradium, and English, French, German, Spanish and Portuguese are all supported from day one.

Alongside the English model, there are dedicated 40M models for German, French, Spanish and Portuguese, plus a multilingual one that handles all five. Dozens of voices are available, selected via a parameter. There is a catch, though: only the English Phonon is released under Apache 2.0, while the other language models come under CC BY-NC 4.0 and are therefore limited to non-commercial use. Anyone planning to build a non-English read-aloud feature into a product needs to take a closer look.

The most practical part is how easily it can be chained. The output of a chat model can be piped straight into Phonon, and because text is spoken as soon as it arrives, playback starts after the first sentence. Developers address the models via the OpenAI-compatible /v1/audio/speech endpoint – existing code only needs a new base URL. Output is 24 kHz mono, WAV by default.

Under the hood, Ollama uses Apple's MLX framework. On a MacBook Pro with M5 Max, Ollama claims to be 35% faster than Gradium's own runtime, with the same word error rate as the reference implementation in Hugging Face Transformers. These are the developer's own figures, and the comparison is not entirely like for like: Ollama runs in bf16, while Gradium's runtime partly uses 8-bit weights.

Windows and Linux are supposed to follow "soon", but Ollama has not given a date. For Mac users who would rather not send their data to the cloud anyway, it is still a nice step: a chatbot that thinks locally and reads aloud locally used to be a tinkering project – now it is a single command in the terminal.