Ollama Phonon Explained: Free Local Text-to-Speech on Your Mac (2026 Guide)

For years, turning text into speech on a computer meant sending your words to someone else’s server. You typed a sentence, an API in a distant data centre read it aloud, and a bill ticked upward with every character. Ollama Phonon breaks that pattern completely. Released with Ollama 0.40.11 on October 24, 2026, it lets an Apple Silicon Mac speak entirely on its own hardware — no internet connection, no subscription, and no per-character charge. In this guide, we explain what Phonon is, how to install and use it, which voices and languages it supports, what it costs (nothing, with one important catch), and how it compares with cloud text-to-speech services such as ElevenLabs and AssemblyAI.

Table of Contents

Ollama Phonon running on a MacBook on a desk

What Is Ollama Phonon?

Ollama Phonon is a family of small, open text-to-speech models built by Gradium, a French speech-technology startup, and now distributed through Ollama, the popular free app for running AI models locally on your own computer. Until this release, Ollama was almost entirely a way to run chatbots offline. You could download a language model, ask it questions, and get answers without anything leaving your machine. Speech was the missing piece. Phonon fills that gap: it is the first speech model Ollama has ever supported, and it turns the app from a text-only tool into a complete local voice stack.

The idea is simple. You give Phonon a piece of text, and it reads it aloud on your Mac. Because the model runs on your own hardware, there is no API key to manage, no account to create, and no usage meter running in the background. The default English model is tiny by modern AI standards — just 31 million parameters — which is why it can run comfortably on an ordinary Apple Silicon laptop rather than demanding a server-grade graphics card.

What makes Phonon feel different from older offline speech tools is streaming. The audio starts playing while the rest of the sentence is still being generated, so a long reply from a chatbot can begin speaking after the first sentence or two instead of making you wait for the whole paragraph. If you pipe a chat model’s output straight into Phonon, the spoken answer follows the text almost immediately, which makes a local AI assistant feel far more like a conversation and far less like a dictation machine.

Ollama Phonon at a Glance

  • What it is: a family of small text-to-speech models from Gradium, distributed through Ollama
  • Launched: October 24, 2026, with Ollama version 0.40.11
  • Platform: Apple Silicon Macs only at launch; Windows and Linux support is promised but undated
  • Languages: English, French, German, Spanish, and Portuguese, plus a multilingual model
  • Voices: dozens of built-in voices, including Zoey, Easton, and Freya
  • Price: free to download and run; no subscription and no per-character billing
  • Licence: the default English model is Apache 2.0 (commercial use allowed); the other language models are CC BY-NC 4.0 (non-commercial use only)
  • Audio output: 24 kHz mono WAV by default, with raw PCM also available
  • Developer access: an OpenAI-compatible /v1/audio/speech endpoint, so existing code needs only a new base URL

How Ollama Phonon Works

Under the hood, Ollama Phonon relies on Apple’s MLX framework, an open-source machine-learning toolkit designed for Apple Silicon’s unified memory architecture. Instead of copying data back and forth between separate processor and graphics memory, MLX lets the model work directly with the memory shared by the whole chip. That is a large part of why such a small model can speak so quickly on a MacBook.

Ollama reports that, on a MacBook Pro with an M5 Max chip, its MLX-based version of Phonon generates speech about 35% faster than Gradium’s own reference runtime, while matching the reference implementation’s word error rate in Hugging Face Transformers tests. Those are Ollama’s own figures, not independent benchmarks, but they point in a useful direction: this is not a slow, compromised offline fallback. For short-form narration, assistant replies, and accessibility reading, the speed is genuinely practical.

The workflow has three stages. First, the model converts your text into a sequence of speech sounds. Second, it renders those sounds into audio samples at 24 kHz, which is perfectly clear for spoken voice even though it is below the 44.1 kHz standard used for music. Third, because generation is streamed, the first chunk of audio can play while later chunks are still being produced. The result is a voice that starts almost immediately and keeps pace with the text.

One technical detail is worth knowing: Phonon runs in bf16 precision, while Gradium’s own runtime partly uses 8-bit weights. In plain terms, Ollama trades a little memory for accuracy, which helps explain why its output quality matches the reference version rather than falling behind it.

How to Install Ollama Phonon on Your Mac

Getting started with Ollama Phonon takes about five minutes if you already use Ollama, and only slightly longer if you are installing it for the first time. You need a Mac with Apple Silicon (any M1, M2, M3, M4, or M5 machine) and Ollama version 0.40.11 or later.

Step 1: Update or Install Ollama

Download Ollama from its official website if you do not have it yet, or update your existing installation to version 0.40.11 or newer. Speech support arrived in that release, so older versions will not recognise speech models at all. You can check your version by running ollama –version in Terminal.

Step 2: Pull the Phonon Model

Open Terminal and download the default English model with a single command:

ollama pull phonon

This fetches Gradium’s 31-million-parameter English checkpoint, which takes roughly 121 MB of memory when running. If you want a different language, pull that model’s name instead — for example, the French, German, Spanish, Portuguese, or multilingual checkpoints. Each one is a separate small download.

Step 3: Speak Your First Line

Once the model is downloaded, ask it to speak with another single command:

ollama run phonon “Hello from Ollama.”

On a Mac with speakers, the line plays straight through the built-in audio using afplay. If you run the same command on a headless Mac mini with no speakers attached, or redirect the output to a file, Phonon writes a WAV file instead of playing sound. Run ollama run phonon on its own and you get an interactive prompt where every line you type is spoken when you press Enter.

Ollama Phonon developer working on a laptop

Step 4: Chain It With a Chat Model

The most interesting setup pipes a chat model’s output directly into Ollama Phonon. Ask your local language model a question, send its answer into the speech model, and the reply is read aloud as it streams in. There is no waiting for the full answer to finish generating, and no audio file to manage in between. For anyone building a local voice assistant, this two-model chain — one model thinking, one model speaking — is the whole point of the release.

Choosing a Voice in Ollama Phonon

Ollama Phonon ships with a catalogue of voices rather than a single default speaker. Gradium’s voice page lists 102 voices for the small Phonon model: 90 catalogue voices across seven locales, plus 12 English character voices. In practice, most users will start with the well-known names such as Zoey, Easton, and Freya, and then explore from there.

Selecting a voice is a one-flag change. To hear a specific voice, add the voice option to your command:

ollama run phonon –voice Zoey “This is Phonon speaking.”

If you do not specify a voice, Phonon uses the checkpoint’s built-in default voice. To see every voice available for the model you downloaded, run ollama show phonon, which lists them all. Voice names work regardless of capitalisation, and accented characters are handled gracefully, so you do not need to worry about exact spelling.

Voice choice matters more than it first appears. A calm, neutral voice suits long-form reading and accessibility use, while a livelier voice fits a demo or a character in a project. Because switching voices costs nothing and takes seconds, it is worth trying several before settling on one for a regular task such as reading articles aloud or narrating a daily briefing.

Ollama Phonon person listening with headphones

Using Ollama Phonon in Your Own Apps

For developers, the most important feature of Ollama Phonon is its OpenAI-compatible endpoint. Any application that already talks to a cloud speech API can point at a local Ollama server instead, usually by changing nothing more than the base URL. The endpoint is /v1/audio/speech on your local Ollama host, which by default runs on port 11434.

A basic request looks like this:

curl http://localhost:11434/v1/audio/speech -d ‘{“model”: “phonon”, “input”: “Hello from Phonon.”, “voice”: “Zoey”, “response_format”: “wav”}’

The default response format is WAV: 16-bit, mono, 24 kHz audio. If you need raw samples without a file header, ask for PCM instead, which returns raw 16-bit little-endian samples at the same rate. Speech-only models reject chat, completion, and embedding calls with an error, which is expected behaviour — Phonon speaks, and nothing else.

This compatibility has a practical consequence. Tutorials, libraries, and tools built for OpenAI’s speech API can be tested against a free local model before a single dollar is spent on cloud usage. A developer prototyping a voice feature can iterate locally with Phonon, then switch to a paid cloud voice only if the project outgrows what a small local model can do. That workflow — local first, cloud when justified — is exactly the audience Ollama serves.

Ollama Phonon studio microphone for voice recording

Ollama Phonon Languages and Models

Ollama Phonon supports five languages from day one: English, French, German, Spanish, and Portuguese. Each language has its own dedicated 40-million-parameter model, except for the default English checkpoint, which is the smaller 31-million-parameter version. A multilingual model covering all five languages is also available for projects that switch between languages without swapping models.

The language split matters because of the licence difference described below, but it also matters for quality. Dedicated per-language models are trained specifically for one language’s sounds and rhythms, which generally produces more natural pronunciation than a single model stretched across many languages. If your project only needs English, the default model is the obvious choice. If you need German or French narration, use that language’s dedicated model rather than the multilingual one.

Coverage beyond these five languages is the obvious gap. There is no Hindi, Urdu, Arabic, Japanese, or Chinese model in the Phonon family yet, and Gradium has not announced a timetable for more languages. For English and Western European language users, Phonon is immediately useful. For everyone else, it is a promising preview rather than a finished tool.

The Licence Catch You Must Know

Here is the detail that catches many people out, and it is the single most important paragraph in this guide. Not every Ollama Phonon model carries the same licence, and using the wrong one in a commercial project could create a legal problem.

Ollama Phonon English Model: Apache 2.0

The default English model, Phonon 31M, is released under the Apache 2.0 licence, which allows commercial use. Gradium’s voice catalogue is also Apache 2.0. If you build a paid product, a client project, or a monetised video narration pipeline around the English model, the licence permits it. This is the model most readers should start with, and it is the only one you need for English-only work.

The Other Language Models: CC BY-NC 4.0

The 40-million-parameter English, French, German, Spanish, Portuguese, and multilingual checkpoints are released under CC BY-NC 4.0, which limits use to non-commercial purposes. In plain terms, you can use them for personal projects, research, education, and experimentation, but not in a product or service you charge for. Ollama’s own library page states this explicitly, so there is no ambiguity — but it is easy to miss if you simply pull a model and start building.

The practical rule is simple: for anything commercial, stick to the default English Phonon model and its Apache 2.0 voices. For a personal French narration project or a classroom Spanish demo, the non-commercial models are fine. When in doubt, check the model card for the exact checkpoint you downloaded before shipping anything.

Ollama Phonon vs Cloud Text-to-Speech

How does a free local model compare with the cloud services most people actually use? The honest answer is that Ollama Phonon wins on price and privacy, and loses on voice quality, language coverage, and advanced features. Which side of that trade-off matters depends entirely on your use case.

  • Price: Ollama Phonon costs nothing beyond the Mac you already own. ElevenLabs charges per character on paid plans, and AssemblyAI’s TTS-1, launched October 29, 2026, costs $15 per one million input characters. For high-volume or always-on use, the local option saves real money.
  • Privacy: with Ollama Phonon, your text never leaves your machine. Cloud services necessarily receive every word you send them, which matters for confidential documents, unreleased scripts, and internal tools.
  • Voice quality: cloud voices from ElevenLabs and similar providers still sound more expressive and human, with finer control over emotion, pacing, and emphasis. Ollama Phonon is clear and natural, but it is a small model, and it shows on long, dramatic passages.
  • Languages: cloud providers cover dozens of languages; Ollama Phonon covers five. If you need anything beyond English, French, German, Spanish, or Portuguese, the cloud wins today.
  • Features: AssemblyAI’s TTS-1 offers instant voice cloning from a short sample and more than 4,000 voices, while ElevenLabs offers voice design and dubbing tools. Ollama Phonon offers a fixed catalogue of voices and no cloning at all.
  • Latency: cloud services quote figures such as roughly 90 milliseconds to first audio for AssemblyAI’s model, but network round trips add real delay. Ollama Phonon streams locally, so the first sound can arrive almost immediately on a fast Mac.

The fair summary is that Phonon is not a replacement for a premium cloud voice in a commercial production. It is a replacement for paying a cloud bill during development, prototyping, personal use, and any project where “good enough, private, and free” beats “superb, metered, and remote”. Readers who want the full cloud comparison should see our guide to the best AI voice generators of 2026, and our coverage of AssemblyAI TTS-1, the cloud voice model built for AI agents.

Ollama Phonon Performance: Speed and Quality

Speed is where Ollama Phonon genuinely impresses. The combination of a small model, Apple’s MLX framework, and streaming output means speech begins quickly and keeps up with reading pace. Ollama’s claim of 35% faster generation than Gradium’s own runtime on an M5 Max MacBook Pro is a vendor figure, but even allowing for marketing optimism, the experience on an M-series Mac is responsive rather than sluggish.

Quality is best described as good, not spectacular. Spoken sentences are clear, correctly pronounced, and easy to follow, which is exactly what you want for reading articles, narrating a document, or voicing an assistant’s replies. Longer passages can sound a little flat compared with the most expressive cloud voices, and the 24 kHz sample rate, while fine for speech, is not studio-grade audio. For podcasts or commercial narration, a cloud service or a recorded human voice remains the better choice.

Hardware requirements are modest. The English model needs only about 121 MB of memory while running, so even a base-model MacBook Air with 8 GB of unified memory can handle it alongside a small chat model. Larger chat models will compete for the same memory, so on a base machine it is worth keeping the chat model small if you want both thinking and speaking at once. On a Mac with 16 GB or more, the combination runs comfortably.

Ollama Phonon person working from home on a laptop

Who Should Use Ollama Phonon (and Who Should Not)

Ollama Phonon is an excellent fit for several groups. Developers prototyping voice features can iterate for free before committing to a cloud bill. Privacy-conscious users can have documents read aloud without uploading them anywhere. Students and hobbyists get a genuinely useful speech tool at zero cost. Accessibility users gain a free, offline screen-reading voice for their Mac. And anyone building a local AI assistant — the increasingly popular “private ChatGPT on my laptop” project — finally has the speaking half of the setup.

It is a weaker fit for commercial voice production, where cloud voices still sound better and offer cloning, dubbing, and dozens of languages. It is also not yet an option for Windows or Linux users, who must wait for Ollama’s promised ports. And if you need a language outside the five supported ones, Phonon simply cannot help you today. The tool is honest about all of this: it does one job, locally, for free, and it does that job well.

Ollama Phonon Limitations: Windows, Linux, and Missing Features

The largest limitation is platform. Ollama Phonon speech models are available for macOS on Apple Silicon only. Windows and Linux support is described as “coming soon” by Ollama, but no date has been announced, and past experience with Ollama feature rollouts suggests “soon” can mean weeks or months. Windows and Linux users should treat this as a Mac-only story for now and check the Ollama release notes before assuming otherwise.

Other gaps are worth listing plainly. There is no voice cloning: you pick from Gradium’s catalogue rather than creating a voice from a sample. There is no emotion or style control beyond choosing a voice; you cannot ask the model to sound excited, whisper, or emphasise a word. There is no speech-to-text companion in this release — Phonon speaks but does not listen, so a full voice conversation still needs a separate transcription tool. And the non-commercial licence on the non-English models limits what businesses can do with four of the five languages.

None of these gaps makes Phonon a bad tool. They define its lane: a free, private, local voice for Mac users, developers, and personal projects, with commercial English use explicitly allowed. Expect the lane to widen as Gradium adds languages and Ollama ports speech to other platforms.

Privacy: Why Local Speech Matters

Every cloud text-to-speech request sends your text to a company’s servers. For most content that is harmless, but for some it is a genuine concern: an unreleased business plan, a patient’s notes, a legal document, a script under embargo, or simply a private journal entry. Once text leaves your machine, you rely on the provider’s retention policy, security practices, and jurisdiction.

Phonon removes that concern entirely. The model file sits on your disk, the computation happens on your chip, and the audio comes out of your speakers. Nothing is uploaded, logged, or metered. For journalists handling sensitive sources, lawyers reviewing confidential material, or anyone who simply prefers that their computer’s reading habits stay on their computer, that property alone justifies installing it.

There is a second, quieter privacy benefit. Because there is no account, there is no usage history tied to your identity, no billing record of what you had read aloud, and no risk of a future price change or service shutdown cutting off a tool you depend on. The model file you download today keeps working tomorrow, whether or not Gradium or Ollama changes their plans.

Frequently Asked Questions About Ollama Phonon

Is Ollama Phonon really free?

Yes. Ollama Phonon is free to download and run, with no subscription and no per-character billing. The only cost is the Mac you run it on. Remember the licence distinction, though: the default English model allows commercial use, while the other language models are non-commercial only.

Does Ollama Phonon work on Windows or Linux?

Not yet. Speech models in Ollama are available for macOS on Apple Silicon only at launch. Ollama says Windows and Linux support is coming soon, but has not announced a date.

Does Ollama Phonon need an internet connection?

Only to download Ollama and the model files the first time. After that, Ollama Phonon runs entirely offline on your Mac.

How many voices does Ollama Phonon have?

Gradium lists 102 voices for the small Phonon model, including 90 catalogue voices and 12 English character voices. Popular choices include Zoey, Easton, and Freya, and you can list them all with ollama show phonon.

Can Ollama Phonon clone my voice?

No. Ollama Phonon does not offer voice cloning. You choose from its built-in voice catalogue. For cloning, cloud services such as ElevenLabs or AssemblyAI’s TTS-1 are the current options.

What audio quality does Ollama Phonon produce?

Output is 24 kHz mono audio in WAV format by default, with raw PCM also available. That is clear and perfectly suitable for spoken voice, though it is below music-studio quality.

Can I use Ollama Phonon in my own app?

Yes. Ollama Phonon exposes an OpenAI-compatible /v1/audio/speech endpoint on your local machine, so most existing speech-API code works by changing only the base URL.

How is Ollama Phonon different from macOS built-in speech?

macOS has long offered a say command and system voices, and those remain perfectly good for basic reading aloud. Ollama Phonon adds a modern neural voice with streaming, a developer API compatible with OpenAI’s speech endpoint, and a path to chain speech with local language models — the building blocks of a local voice assistant, rather than a single system voice.

Final Verdict: Is Ollama Phonon Worth Installing?

Yes, with the right expectations. Ollama Phonon is the most useful free upgrade to the local AI toolkit since Ollama itself. For Mac users, it takes five minutes to install, costs nothing to run, keeps your text private, and turns a text-only local chatbot into something that can actually talk to you. The streaming playback, the OpenAI-compatible API, and the Apache 2.0 English model make it a serious tool for developers, not just a novelty.

Its limits are real but clearly marked: five languages, Mac only for now, no voice cloning, and a non-commercial licence on everything except the default English model. If you need a premium commercial voice, dozens of languages, or Windows support, the cloud services still win. But if you own an Apple Silicon Mac and have ever wished your local AI could speak, Phonon is the easiest free upgrade available this month — install it, try Zoey or Freya, and hear the difference a local voice makes. For the wider voice landscape, our best AI voice generator guide and our DiffuSpace Clover explainer show where local and cloud AI are heading next.

Sources: 9to5Mac’s report on Ollama’s local text-to-speech launch and Notebookcheck’s coverage of Phonon’s languages and licences.

Leave a Comment

Your email address will not be published. Required fields are marked *