Ollama adds text-to-speech: Phonon runs locally on Macs behind an OpenAI-style speech endpoint
Ollama's October 24 release brings its first speech model, Phonon from Gradium. It runs on Apple Silicon through MLX, with Windows and Linux listed as coming soon. Existing OpenAI SDK speech code works once you change the base URL to localhost.
You need Ollama 0.40.11 or later. ollama run phonon "text" speaks the text, --voice picks a voice from Gradium's catalog (ollama show phonon lists them), and redirecting the output writes a WAV file. Audio is 24 kHz mono, returned as WAV by default or as raw 16-bit PCM.
You can pipe a chat model straight into it, and playback starts after the first sentence instead of waiting for the whole reply. The /v1/audio/speech API streams too, so an app can play audio while it's still being generated.
Licensing splits by model. The English phonon model is Apache 2.0. The 40M French, German, Spanish, Portuguese and multilingual models are CC BY-NC 4.0, which rules out commercial use.
Ollama says it ran 35% faster than Gradium's own runtime on an M5 Max MacBook Pro (45 seconds of speech, median of 10 runs), with the same word error rate as the Transformers reference on 200 Seed-TTS English sentences. Those are Ollama's numbers; we haven't run it.
Why it matters for builders
If you prototype voice agents, you can now test text-to-speech locally for free, with no cloud key, using the same client code you'd ship against OpenAI. Check the license before a non-English voice goes into a product.
Try this
On an Apple Silicon Mac with Ollama 0.40.11 or later, run ollama pull phonon, then ollama run gemma4 --hidethinking "Tell me a fun fact about owls in two sentences." | ollama run phonon. In an app, set your OpenAI client's base_url to http://localhost:11434/v1 and call audio.speech.create with model "phonon".
Sources
Get stories like this every Monday
One tested AI workflow with the prompts, plus 3 news items that matter if you build. Free.
More AI news
- LiteLLM's root-cause report: a master key leaked in May and never revoked led to the poisoned 1.107.0
- Hugging Face releases Carbon-X, a 14.8B open model that reads DNA and answers in English
- Nanbeige4.3-33B: an open 33B model whose card claims it beats Qwen3.8-27B on most agent coding tests
- Anthropic's advice for agent builders: fold most subagents back into one loop with skills