Ollama Adds Text-to-Speech on Mac: Phonon Runs 35% Faster Than Gradium's Native Runtime
Ollama, a tool for running neural networks on your own computer, can now read text aloud. Its first speech model is Phonon from Gradium. On a MacBook Pro with an M5 Max chip, Ollama synthesizes speech 35% faster than Gradium's own runtime, ptts, and the word error rate of the speech matches the reference implementation. That means a local chatbot's replies can be spoken without the cloud, though so far only on Mac and only in English.
AI-processed from Ollama Blog; edited by Hamidun News
Imagine you've launched a chatbot on your laptop. It replies in text, and you want it to speak. Until now, that meant hooking up a cloud speech service or stitching together several separate programs. Now Ollama, a popular tool for running neural networks locally, has added support for speech models. The first is Phonon, a family of small text-to-speech (TTS) models from Gradium.
The headline number of the release is 35%. That is how much faster Ollama is than the closest competitor, Gradium's own runtime called ptts. The benchmark ran on a MacBook Pro with an Apple M5 Max chip: the model generated 45 seconds of speech, and the result is the median of 10 runs. Speed was measured as how many times faster than real time the generation runs: the higher the number, the sooner the audio is ready.
How to Try It
Speech models are currently available only on macOS. Windows and Linux support is promised later. You need Ollama version 0.40.11 or newer. After that, a single command is enough: ollama run phonon "Hello from Ollama." and your computer says the phrase. The voice is selected with the --voice flag (the example uses the Hadley voice from Gradium's catalog), and the output can be saved to a WAV file with ordinary shell output redirection.
Pairing with a chat model is also convenient. The reply can be piped straight into Phonon: the text is spoken as it arrives, so speech starts after the first sentence rather than after the full response is generated. The blog's example combines the gemma4 model with Phonon: the first tells a fact about owls, and the second reads it aloud.
What the Benchmarks Actually Show
On Apple Silicon, Ollama runs through MLX, Apple's machine learning framework. The comparison with Gradium is not entirely apples to apples. Gradium's runtime uses a mix of formats: 8-bit weights on the CPU, fp16 on the Neural Engine, and f32 on the GPU, while Ollama runs the model in bf16. In addition, the numbers come from Ollama's own authors, and there is no independent verification yet.
Quality did not drop with the speedup. Ollama compared the word error rate with the reference implementation in Hugging Face Transformers. The test used 200 English sentences from the Seed-TTS test set, and the generated speech was run through Whisper large-v3 recognition. The lower the error rate, the clearer the speech. Ollama's figure matched the reference: the optimization did not hurt pronunciation.
Integration and Licenses
For practitioners, the key point is that Phonon works through an OpenAI-compatible /v1/audio/speech endpoint. If a project already uses the OpenAI SDK for speech, it is enough to change the base URL to http://localhost:11434/v1 and specify any key, for example "ollama". Examples are given for curl, Python, and JavaScript. The response arrives as a stream, which can be fed directly to a player such as ffplay, so playback starts almost instantly. The audio comes out at 24 kHz, mono: by default as WAV or raw 16-bit samples.
The licensing has a catch. Phonon is released under Apache 2.0, but the 40M-size models fall under CC BY-NC 4.0, meaning non-commercial use only. For a startup, this means you can build a prototype, but you cannot ship a paid product on these weights without a separate agreement. Check the model cards for the terms of the other sizes in the family.
Other limitations: it is Mac-only for now, and all examples and benchmarks are shown for English. No Russian support has been claimed in the material, so do not count on it.
What to Do
If you are building a local voice assistant, a text reader, speech output for internal tools, or a voice-enabled agent prototype on a Mac, it is worth trying now: the entry point is one command and a URL change in existing code. For commercial projects, first check the license of the specific model, and measure the speed on your own hardware.
- Nov 21, 2026NAND flash for SSDs drops 19–32% in spot deals — drives could get cheaper
- Nov 21, 2026Unsloth Updates Compressed Qwen3.8-27B: Accuracy Up 5–10% at the Same Size
- Nov 21, 2026AI Writes Vulnerable Code in 44% of Tests: Positive Technologies Gives Agents a Scanner
- Nov 21, 2026Sber Releases a 14,990-Ruble Speaker with GigaChat That Understands Voice Without Internet
Key Questions Answered
How to Try It?
Speech models are currently available only on macOS. Windows and Linux support is promised later. You need Ollama version 0.40.11 or newer. After that, a single command is enough: ollama run phonon "Hello from Ollama." and your computer says the phrase. The voice is selected with the --voice flag (the example uses the Hadley voice from Gradium's catalog), and the output can be saved to a WAV file with ordinary shell output redirection. Pairing with a chat model is also convenient.
What the Benchmarks Actually Show?
On Apple Silicon, Ollama runs through MLX, Apple's machine learning framework. The comparison with Gradium is not entirely apples to apples. Gradium's runtime uses a mix of formats: 8-bit weights on the CPU, fp16 on the Neural Engine, and f32 on the GPU, while Ollama runs the model in bf16. In addition, the numbers come from Ollama's own authors, and there is no independent verification yet. Quality did not drop with the speedup.
What are the key details on Integration and Licenses?
For practitioners, the key point is that Phonon works through an OpenAI-compatible /v1/audio/speech endpoint. If a project already uses the OpenAI SDK for speech, it is enough to change the base URL to http://localhost:11434/v1 and specify any key, for example "ollama". Examples are given for curl, Python, and JavaScript. The response arrives as a stream, which can be fed directly to a player such as ffplay, so playback starts almost instantly.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.