Ollama adds text-to-speech with Phonon models from Gradium, starting on macOS

October 24, 2026Source: Ollama

Ollama can now turn text into speech, with Phonon as its first speech model. Phonon is a family of small text-to-speech models from Gradium, and support is available on macOS, with Windows and Linux coming soon, according to Ollama.

Key points

Ollama now supports text-to-speech, with Phonon from Gradium as the first speech model.
Speech models run on macOS with Ollama 0.40.11 or later; Windows and Linux support is coming soon.
According to Ollama, it was 35% faster than Gradium's own runtime on a MacBook Pro with M5 Max.
Phonon works with the OpenAI-compatible /v1/audio/speech endpoint and supports English, French, German, Spanish and Portuguese variants.
Phonon is under Apache 2.0, but the 40M models use CC BY-NC 4.0, limiting them to noncommercial use.

Phonon arrives as the first speech model

Ollama, the tool for running AI models locally, announced speech model support on October 24, 2026. The first model is Phonon, which Ollama describes as a family of small, state-of-the-art text-to-speech models from Gradium. Speech models currently run on Apple Silicon Macs, and users need Ollama 0.40.11 or later.

Getting started takes one command to pull the model, ollama pull phonon, followed by a run command such as ollama run phonon "Hello from Ollama."

Voices come from Gradium's voice catalog. Users pick one with the --voice flag, or the voice field in the API, and ollama show phonon lists all available voices.

Languages and licensing

The base phonon model covers English. Other variants are listed by Ollama as follows:

phonon:40m-fr: French
phonon:40m-de: German
phonon:40m-es: Spanish
phonon:40m-pt: Portuguese
phonon:40m-multilingual: all five languages

Phonon is released under Apache 2.0. The 40M models are released under CC BY-NC 4.0, which limits them to noncommercial use.

Speed and accuracy on Apple Silicon

On Apple Silicon, Ollama is accelerated by MLX, Apple's machine learning framework. According to Ollama, on a MacBook Pro with M5 Max it was 35% faster than the next fastest engine, Gradium's own runtime. The measurement generated 45 seconds of speech and used the median of 10 runs.

Ollama notes the two setups differ in numeric precision. Gradium's runtime, ptts, uses 8-bit weights on the CPU, fp16 on the Neural Engine and f32 on the GPU, while Ollama runs the model in bf16.

On accuracy, Ollama says its implementation matches the word error rate of the reference implementation in Hugging Face Transformers. The test used 200 sentences from the Seed-TTS English test set, transcribed with Whisper large-v3.

Reading replies aloud and using the API

Phonon can be combined with a chat model by piping output from one command into the other. Ollama's example runs gemma4 and sends its reply to Phonon. Text is spoken as it arrives, so playback starts after the first sentence. Users can also run ollama run phonon with no text to type lines at a prompt, or redirect the output to save a WAV file.

For developers, Phonon works with the OpenAI-compatible /v1/audio/speech endpoint. Ollama says existing code that uses the OpenAI SDK only needs a new base URL. The announcement includes examples for cURL, Python and JavaScript, all pointing at http://localhost:11434/v1.

The response streams while audio is generated, so it can play right away. Audio is 24 kHz mono, returned as WAV by default or as raw 16-bit samples when response_format is set to pcm.

What it means for users

For Mac users, speech generation can now run through the same local tool they may already use for language models, with no separate runtime. The OpenAI-compatible endpoint means apps built against that API can be pointed at a local server. Windows and Linux users will have to wait, as Ollama only says support is coming soon. Anyone planning commercial use should check the licenses, since the 40M models are limited to noncommercial use.