Ollama Brings Local Text-to-Speech to Mac with Phonon Models
As of October 24 2026 macOS users can run Phonon text-to-speech models locally in Ollama 0.40.11 turning chat replies into spoken audio that streams in real time. The models…
- Phonon family of small TTS models now available in Ollama for Apple Silicon Macs with Windows and Linux support coming soon
- Multiple voices and languages supported; main model under Apache 2.0 40M variants limited to noncommercial use under CC BY-NC 4.0
- Streams audio as it generates and can pipe output from any local chat model for immediate spoken replies
Local TTS now available in Ollama
Ollama added text-to-speech support on October 24 2026. The first models are the Phonon family from Gradium. macOS users with Apple Silicon can run them immediately while Windows and Linux versions are still in development.
On an M5 Max MacBook Pro Ollama delivered 35 percent higher throughput than Gradium's own runtime while using bf16 precision and matching the reference word error rate on a 200-sentence English test set transcribed by Whisper large-v3. The command ollama run phonon "Hello from Ollama." produces audio right away.
Voices languages and real-time use
Users select voices from Gradium's catalog with the --voice flag or the API parameter. The base phonon model supports English; the smaller 40M models add French German Spanish Portuguese and a multilingual version.
A practical feature is piping output from a local chat model directly into Phonon. For example ollama run gemma4 --hidethinking followed by a pipe to phonon speaks the reply as soon as the first sentence arrives. Output can be saved as a WAV file or streamed as raw 16-bit samples at 24 kHz mono.
API compatibility and getting started
Phonon implements the OpenAI-compatible /v1/audio/speech endpoint so code using the official OpenAI SDK only needs a changed base URL to run locally. Responses can stream while audio is generated allowing immediate playback.
To start download Ollama 0.40.11 or later pull the model with ollama pull phonon and use ollama show phonon to list available voices. Full details are in the model page speech documentation and OpenAI compatibility notes.
Licensing
The primary phonon model is released under Apache 2.0. The 40M variants use the CC BY-NC 4.0 license which restricts them to noncommercial use.
Why it matters for uncensored AI
Our analysis starts from the source fact that Ollama runs these speech models entirely on the user's own hardware with local chat model integration. This development gives people another building block for keeping both conversation and spoken output under their direct control instead of sending text to remote services.
For users who prioritize private local workflows the addition of on-device TTS removes one more reason to rely on hosted platforms. That practical gain in ownership and data locality supports the broader choice between local tools and cloud services.
