Ollama Text-to-Speech: How to Run Phonon Voices on a Mac

Ollama can now talk back. On October 24, 2026, the popular local AI runtime added text-to-speech, and its first speech model is Phonon, a family of tiny open models from the voice AI company Gradium. If you already use Ollama to run language models on your own machine, you can now give those models a voice without sending a single word to the cloud. Here is what Ollama text-to-speech supports today, how to try it, and what the licences allow.

• Requirement: Ollama 0.40.11 or later on an Apple Silicon Mac; Ollama says Windows and Linux support is coming soon
• Model size: Gradium says Phonon uses 31 million parameters in its English model, versus 82 million for the popular Kokoro TTS model
• Speed: Ollama measured its MLX-accelerated build at 35% faster than Gradium's own runtime on a MacBook Pro with M5 Max
• Languages: English plus French, German, Spanish and Portuguese, with a single multilingual 40M model covering all five

What Is Phonon, the New Ollama Speech Model?

Gradium, founded by the cofounders of the Kyutai research lab, released Phonon on October 21 at NVIDIA GTC in Berlin. It is designed to generate speech entirely on the user's hardware, which keeps both the text and the audio on the device and removes network latency. Gradium's announcement describes 31M and 40M parameter variants, sized to ship inside a mobile app or load in a browser tab, with a memory footprint quoted between 64.5 MB and 90 MB on the same page. The company says Phonon won overall preference in blind human evaluations against Kokoro, Magpie and Pocket TTS; that is Gradium's own comparison, so treat it as a vendor claim until independent listening tests appear.

Ollama was one of the named launch partners, alongside Hugging Face, Arm, Qualcomm, LM Studio and MLX Audio. Three days later, Phonon arrived in the Ollama library.

How Do You Run Text-to-Speech in Ollama?

Getting started takes two commands. Pull the model with ollama pull phonon, then speak a line with ollama run phonon followed by your text in quotes. A few useful patterns from Ollama's announcement:

• Pick a voice with the --voice flag (for example Hadley or Zoey); ollama show phonon lists every available voice
• Read a chat reply aloud by piping a language model into Phonon — speech starts after the first sentence arrives, so you are not waiting for the whole answer
• Save to a file by redirecting the output to a WAV file
• Use the API through the OpenAI-compatible /v1/audio/speech endpoint on localhost port 11434, so existing OpenAI SDK code only needs a new base URL

Audio comes back as 24 kHz mono, either WAV by default or raw 16-bit PCM, and it streams as it is generated so playback can begin immediately.

How Fast and Accurate Is Ollama's Phonon?

Ollama runs Phonon in bf16 through MLX, Apple's machine learning framework. In its tests generating 45 seconds of speech on an M5 Max, the median of 10 runs came out 35% faster than Gradium's ptts runtime. Ollama also checked accuracy: on 200 English sentences from the Seed-TTS test set, transcribed back with Whisper large-v3, its word error rate matched the Hugging Face Transformers reference implementation. Neither post publishes absolute real-time factors, so the useful takeaway is relative: the Ollama build is quicker without trading away intelligibility.

On the GPU side, Gradium says its CUDA backend runs 8x faster than CPU and produces the first audio in 5 ms on an NVIDIA L40S, which bodes well for the promised Windows and Linux versions.

What Licence Does Phonon Use?

This is the detail worth reading twice. The English phonon model is released under Apache 2.0, so it is fine for commercial products. The 40M language models — French, German, Spanish, Portuguese and the multilingual build — use CC BY-NC 4.0, which limits them to noncommercial use. Voice cloning and voice design are not part of the open release; Gradium offers those through its paid API.

Why Local Text-to-Speech Matters

Speech is the missing piece for fully offline assistants, accessibility tools and reading apps. Ollama already pairs well with local speech recognition such as IBM's Granite Speech 5 models, and Apple Silicon owners have seen steady gains like the Gemma 4 MLX speedup in Ollama. With Phonon, a private listen-think-speak loop on a laptop is now a few terminal commands away.

Explore more open and local models in our AI coverage.

Sources: Ollama — Speech models in Ollama — October 24, 2026; Gradium — Gradium releases Phonon — October 21, 2026.