Ollama Phonon on Mac: How to Generate AI Speech and Save WAV Audio
Imagine asking a local AI assistant a question and hearing its answer through your Mac’s speakers instead of reading another screen. That is what Ollama Phonon makes possible.
Ollama Phonon turns written text into spoken audio. It can also read a local AI model’s response aloud as the text arrives. Ollama introduced speech support on October 24, 2026, for Apple Silicon Macs, recommending Ollama version 0.40.11 or later.
The official Ollama speech announcement provides the core commands. This guide explains how to set up Phonon, generate a short sentence, save a WAV file, connect a local AI model, and check model licensing before using the audio.
However, Phonon is the speech component—not the AI model that writes your answer. If you want an assistant to generate a response and then speak it, you need a separate text-generation model.
Editorial note: This is a documentation-based guide checked against official Ollama sources on October 27, 2026, with current model details reviewed on November 5, 2026. Tech4GenZ has not executed these steps on an Apple Silicon Mac or evaluated the resulting audio. Commands follow official examples or clearly identified shell adaptations; they are not reported hands-on test results.
In this article
- Ollama Phonon on Mac: How to Generate AI Speech and Save WAV Audio
- Key Takeaways
- Which Macs Support Ollama Phonon Text-to-Speech?
- Which Ollama Phonon Model and License Should You Choose?
- Set Up Ollama Phonon and Generate Your First Sentence
- How Do You Save Ollama Phonon Audio as a WAV File?
- How Do You Make Ollama Read an AI Reply Aloud?
- What Stays Local in an Ollama Phonon Workflow?
- Ollama Phonon Troubleshooting: Checks Before You Retry
- Frequently Asked Questions
- Turn Your First Spoken Reply Into a Usable Result
Key Takeaways
- Use an Apple Silicon Mac and update Ollama first. This guide follows the launch recommendation of version 0.40.11 or later.
- Start with
phonon. The default English checkpoint is Apache 2.0 licensed, while the 40M checkpoints use a noncommercial license. - Test one sentence before connecting another model. Then save a WAV file and listen to it separately.
- Check the installed voice list. A voice appearing in a catalog does not prove that it is available in your downloaded model.
- Keep both generation stages local if that is your goal. Initial model downloads still require an internet connection, and cloud-model variants change the workflow.
- Listen before publishing. Fluent speech does not prove that an AI answer is factually correct or that every name was pronounced properly.
Which Macs Support Ollama Phonon Text-to-Speech?
First, open Apple menu → About This Mac and check the chip information. This walkthrough is intended for Apple Silicon Macs, not Intel Macs.
Next, install or update Ollama using the official download page. Open the application and launch Terminal.
The following availability snapshot follows Ollama’s October 24 speech announcement, rather than assuming that general Ollama compatibility means speech-model compatibility.
| Requirement | What to use for this guide |
|---|---|
| Hardware | Apple Silicon Mac |
| Ollama version | 0.40.11 or later, following the launch recommendation |
| Speech model | Default phonon English model |
| Other operating systems | Windows and Linux speech support announced as coming soon |
| Extra model for spoken AI replies | A separately downloaded local text model that fits your Mac |
| Internet connection | Required for initial software and model downloads |
Run these checks in Terminal:
uname -m
ollama --version
In a native Apple Silicon Terminal session, uname -m should report arm64. However, an x86_64 result can also indicate that Terminal is running under Rosetta.
Therefore, verify the chip in About This Mac rather than declaring your hardware unsupported based on that command alone.
Why use Ollama 0.40.11 instead of 0.40.9?
The Ollama speech documentation and Phonon model readme list version 0.40.9 or later. However, the launch announcement recommends 0.40.11 or later.
For beginners, this guide uses the more conservative launch recommendation. It does not claim that 0.40.11 was the first version with any speech support.
Which Ollama Phonon Model and License Should You Choose?
For the first English example, use phonon without a suffix.
According to the Phonon model card, the default model uses a 31M English checkpoint released under Apache 2.0. By contrast, the 40M checkpoints use CC BY-NC 4.0.
A model tag is the part after the colon in a model name. Consequently, changing phonon to phonon:40m-en is not merely a cosmetic naming change. It selects a different checkpoint with different licensing terms.
| Model choice | Documented language | License |
|---|---|---|
phonon |
English, 31M checkpoint | Apache 2.0 |
phonon:40m-en |
English | CC BY-NC 4.0, noncommercial |
phonon:40m-fr |
French | CC BY-NC 4.0, noncommercial |
phonon:40m-de |
German | CC BY-NC 4.0, noncommercial |
phonon:40m-es |
Spanish | CC BY-NC 4.0, noncommercial |
phonon:40m-pt |
Portuguese | CC BY-NC 4.0, noncommercial |
phonon:40m-multilingual |
English, French, German, Spanish, and Portuguese | CC BY-NC 4.0, noncommercial |
Model tags can change licensing conditions. This diagram summarizes the linked model card.
Which model is better for commercial use?
For a potentially commercial English workflow, start by evaluating the Apache-licensed default model and checking the licenses of any voice assets you use.
Do not substitute a 40M tag for an ad-supported demo or client project without resolving its noncommercial restriction.
The Apache 2.0 license does not impose a noncommercial-only restriction, but it includes conditions, including redistribution obligations.
Meanwhile, the CC BY-NC 4.0 license prohibits commercial use and requires attribution when sharing licensed material.
Neither license should be treated as a blanket guarantee covering every script, voice, person’s likeness, or intended use of generated audio.
Also, do not interpret “multilingual” as “every language.” The published list above does not establish Urdu support.
Set Up Ollama Phonon and Generate Your First Sentence
Start small so that a failure has only one likely source.
First, download the default speech model:
ollama pull phonon
Then inspect it:
ollama show phonon
Keep the exact model name and displayed capabilities or voices in your notes. Because tags can be updated, save the model ID shown by ollama list when recording a reproducible test.
Next, use the official first-sentence command:
ollama run phonon "Hello from Ollama."
Expected behavior—not a reported Tech4GenZ test result: The command should play speech through your computer’s selected audio output.
If it fails, stop here and use the troubleshooting checks below. Adding a chat model will not fix a speech-only failure.
Alternatively, start an interactive session:
ollama run phonon
Type a short line and press Enter.
The Ollama speech guide documents Ctrl+C to stop speaking and /bye or Ctrl+D to exit the session.
Check US and UK Voice Availability Before Choosing an Accent
The Gradium voice catalog lists Zoey under US English and Freya under UK English.
However, the Ollama sources are not fully aligned. The Phonon model readme describes a default voice, while the speech documentation describes additional named voices.
Therefore, treat ollama show phonon as your first check of the installed package.
Use the following examples only if those names appear in its voice list:
ollama run phonon --voice Zoey "Maya will send the report on November fourteenth."
ollama run phonon --voice Freya "Oliver will send the report on the fourteenth of November."
These are fictional listening scripts, not verified audio samples.
If either voice is absent, omit --voice and finish the default-voice test instead. Do not assume that the same voice name works across every Phonon tag.
Also, changing the date format is not an accent control. It simply gives the speech model wording that may be easier to pronounce naturally.
How Do You Save Ollama Phonon Audio as a WAV File?
Once direct playback works, redirect the output to a file.
The documented WAV workflow changes the destination from your speakers to a recording:
ollama run phonon "Save this to a file." > phonon-first-reply.wav
The file is written in your current Terminal directory.
To see that location and inspect the result on your Mac, run:
pwd
ls -lh phonon-first-reply.wav
file phonon-first-reply.wav
afplay phonon-first-reply.wav
Here is what each command does:
pwdshows the current folder.lsshows the file and its size.fileinspects the file format.afplayplays the saved recording on macOS.
These inspection commands are shell additions to the official save example, not evidence that this article’s file has already been generated.
Important: The > operator overwrites an existing file with the same name. Therefore, use a new filename when keeping different attempts.
Ollama documents WAV output as 16-bit, mono, 24 kHz. It also notes that streamed WAV headers use maximum values for size fields because the final length is not known when the header is written.
Consequently, an unusual duration display alone is not proof of a broken recording. Listen to the file and inspect it in your intended editor before delivery.
How Do You Make Ollama Read an AI Reply Aloud?
You need two models for this step:
- A text model that writes the answer.
- Ollama Phonon, which speaks the answer.
The vertical bar, |, passes the first command’s text output into the second command.
Ollama’s official reply-piping example uses gemma4. However, a text model is a separate download and can be much larger than Phonon.
Check the Gemma 4 model listings and choose a local variant that fits your Mac before proceeding. Do not pick a cloud tag if you intend to keep answer generation local.
If you choose the same model name as the official example, download it first:
ollama pull gemma4
Then run the short-response workflow:
ollama run gemma4 --hidethinking "Tell me a fun fact about owls in two sentences." | ollama run phonon
The speech documentation says playback can begin after the first sentence or two arrives. However, that is a description of streaming behavior, not a measured latency promise for your Mac.
Model loading, answer generation, and speech synthesis all affect the wait.
For a speech-friendly result, ask the writing model for short prose without code, tables, or lists. Also, --hidethinking keeps displayed reasoning out of the pipe; it does not fact-check the answer.
Save the Spoken AI Reply Instead of Playing It Immediately
Suggested review workflow, not a record of completed testing.
After the direct pipe works, combine the same pipeline with WAV redirection:
ollama run gemma4 --hidethinking "Explain what a checksum does in two short sentences. Use plain prose, not Markdown." | ollama run phonon > phonon-checksum-reply.wav
This combines Ollama’s documented piping and saving patterns. The exact combined example has not been executed for this guide.
For factual material, a more cautious workflow is to generate and review the text first:
ollama run gemma4 --hidethinking "Explain what a checksum does in two short sentences. Use plain prose, not Markdown." > reply.txt
Open reply.txt, check the explanation against a trustworthy source, and edit it if needed.
Only then turn that approved text into audio:
ollama run phonon < reply.txt > phonon-reviewed-reply.wav
afplay phonon-reviewed-reply.wav
Unlike the immediate pipe, this staged version gives you a review point before the answer becomes narration.
Nevertheless, a reviewed script still needs a listening check for pronunciation, missing words, and repeated phrases.
What Stays Local in an Ollama Phonon Workflow?
In the intended setup, your downloaded local text model writes the answer, the downloaded Phonon model synthesizes speech, and your Mac plays or saves it.
However, the initial Ollama installer and model pulls require downloads.
Ollama’s privacy and local-versus-cloud FAQ says it does not see prompts or data when you run locally. That statement should not be stretched to cover cloud-hosted models, a remote Ollama server, third-party integrations, or files synchronized by other software.
Before handling sensitive material, verify which server and model you are using.
In addition, remember that saved WAV files, text files, shell history, and screen recordings can retain content even when inference is local.
For the first run, use the harmless sample sentences instead of private work.
This is also a different task from choosing a hosted narration API. If that is your actual goal, the Tech4GenZ HeyGen Voice Turbo guide covers that separate workflow’s model-selection, capacity, and permission checks. It is related reading, not a substitute for this local Mac setup.
Ollama Phonon Troubleshooting: Checks Before You Retry
First, isolate the failing stage.
If a fixed sentence will not speak, investigate speech setup. Conversely, if fixed text works but a generated reply does not, investigate the text model and pipe before changing voices.
| Symptom | First check | Next action |
|---|---|---|
ollama is not found |
Installation and Terminal command availability | Reopen Ollama and Terminal; use the official installer’s setup instructions |
| Cannot connect to Ollama | Whether the Ollama app is running | Open the app, then retry the single-sentence example |
| Unsupported platform or model-load failure | Apple chip, version, and exact tag | Use the supported Mac setup; update Ollama and inspect the error |
| No sound but no obvious model error | Selected output device and volume | Save a WAV and try afplay to separate generation from playback |
| Unknown voice | ollama show phonon for the installed tag |
Remove --voice, then choose only a listed name |
| Empty or invalid recording | Terminal errors and file size/type | Rerun a short fixed sentence; keep error messages separate from audio |
| Pipe hangs or the Mac is under memory pressure | Whether the text model works by itself | Try a fitting local model and close unnecessary applications |
| Speech reads awkward symbols or dates | The exact text sent to Phonon | Use clean prose and spell dates or ambiguous abbreviations out |
Do not redirect errors into the audio file with 2>&1. That shell instruction merges error text with the output stream and can spoil a recording.
Instead, read the error in Terminal and keep the WAV destination separate.
FAQ: Does Phonon Accept the Normal Chat API?
No. The speech API documentation says speech models reject chat, completion, and embedding requests.
Use the documented /v1/audio/speech endpoint for an API integration.
Some generic examples on the model-library page show chat calls, so use the speech-specific guide for this task.
FAQ: Can I Request MP3, Faster Speech, or Pronunciation Instructions?
Do not assume those options are supported.
The speech guide documents WAV and raw PCM, and says unsupported formats, speeds, or instructions return a 400 error rather than being silently ignored.
Therefore, begin with the documented defaults and edit the input text when a name or number needs clarification.
FAQ: How Long Can the Input Be?
The documented speech API limit is 4,096 characters per request. However, Ollama says the CLI splits longer text at paragraph and sentence boundaries.
For a first spoken reply, use one or two short sentences rather than making a long document your first test.
These troubleshooting answers were checked against the speech documentation on October 27, 2026. Recheck the official instructions after an Ollama update, particularly if platform availability or supported controls change.
Frequently Asked Questions
Which Macs can run Ollama Phonon?
Ollama Phonon is for Apple Silicon Macs. Check your chip under Apple menu, then About This Mac, and use Ollama 0.40.11 or later.
Which Phonon model should I start with?
Start with the default phonon English model. It is Apache 2.0 licensed, while the 40M model tags use a CC BY-NC 4.0 noncommercial licence.
How do I save Phonon audio as a WAV file?
Redirect the output to a file, for example ollama run phonon "Save this to a file." > phonon-first-reply.wav. Ollama documents WAV output as 16-bit, mono, 24 kHz.
What do I need to make Ollama read an AI reply aloud?
You need two models: a text model that writes the answer and Ollama Phonon, which speaks it. For example, pipe the text model’s output into Phonon.
Does an Ollama Phonon workflow stay local?
In the intended setup, your downloaded local text model writes the answer and the downloaded Phonon model speaks it. The initial Ollama installer and model downloads still require an internet connection.
Turn Your First Spoken Reply Into a Usable Result
Start with one sentence, save a WAV, and only then connect your local assistant.
That sequence makes Ollama Phonon easier to diagnose without hiding failures inside a larger workflow.
If you later pair approved narration with a visual explainer, keep the spoken wording and on-screen text consistent. For that separate design stage, the Canva Craft editing checks offer a useful revision checklist. They do not establish that a particular Canva workflow supports your audio file.
Finally, listen against the script, verify the selected model and voice licenses, and retain the exact commands with your approved recording.
A successful first run is useful. A recording you can explain and reproduce is the better outcome.





[…] distinguish live interpretation from speech generation. The Ollama Phonon Mac voice guide covers reading written AI replies aloud. By contrast, that is a separate workflow and does not […]
[…] See also: Ollama Phonon on Mac: Read AI Replies Aloud […]