A local AI assistant can draft an answer, explain a document or help with code. Having that answer spoken aloud usually requires another component.
Ollama now integrates that component directly. On October 24, 2026, it announced support for speech models, starting with Phonon, a family of compact text-to-speech models from Gradium.
The integration initially supports Apple Silicon Macs. Ollama says Windows and Linux support is coming, without providing a release date in the announcement. (Ollama’s speech models announcement)
For users and developers, the useful question is what becomes easier when speech generation joins an existing local AI workflow.
Availability and documentation checked on October 27, 2026.
What Phonon actually does
Phonon converts written text into spoken audio. You provide the words; the model generates their pronunciation and delivery.
That makes it suitable for reading an assistant’s response, narrating a passage or adding spoken output to an application. Ollama can stream the audio during generation, allowing playback to begin before the complete passage is ready. (Ollama speech documentation)
A conversational voice assistant still needs other components: something to capture and understand speech, a system to decide what to say, and logic to manage the conversation.
Phonon supplies the speaking stage.
Small models, with an important distinction about memory
Gradium designed Phonon for execution on users’ devices. Its model cards describe compact checkpoints that can run on hardware ranging from laptop CPUs to phones and browsers.
Two examples illustrate the size of the underlying models:
| Checkpoint | Parameters, excluding the audio tokenizer | Quantized weights file |
|---|---|---|
| Phonon 31M English | 31 million | 64.5 MB |
| Phonon 40M French | 40 million | 74.6 MB |
These figures describe Gradium’s eight-bit quantized weights files. They are not the total memory required by a running application, nor the download sizes of every Ollama package. (English model card, French model card)
This distinction matters when comparing speech models with the much larger language models used to generate answers. A compact speech component may be practical on a device even when the conversational model needs more resources.
Our guide to local AI memory requirements explains why download size, system RAM and graphics memory should be considered separately.
What developers can build with it
The immediate opportunity is to add spoken output to tools that already produce useful text.
Consider a local document assistant. It could retrieve a passage, generate a short explanation and pass that explanation to Phonon for playback.
Other possible uses include:
- Reading selected passages while a user works in another application.
- Producing draft narration for a prototype.
- Giving spoken feedback in an educational tool or interactive experience.
These are illustrative applications, rather than measured productivity results. Their usefulness depends on pronunciation, listening comfort and how much control the interface gives the user.
A good implementation should make it easy to stop playback, repeat a sentence and return to the original text.
A simple first test in Ollama
On a supported Mac, update Ollama to a current release and download the model:
ollama pull phononThen ask it to speak a short sentence:
ollama run phonon "Your document summary is ready."To save the generated audio instead of playing it immediately:
ollama run phonon "Your document summary is ready." > summary.wavOllama also documents an OpenAI-compatible /v1/audio/speech endpoint for applications. It returns WAV audio by default and supports raw PCM output. (Ollama speech documentation)
For an initial evaluation, use text containing names, dates, abbreviations and numbers that appear in your actual documents. A pleasant demonstration sentence tells you less than a representative sample.
French support comes with different licensing conditions
The default English checkpoint is only one member of the Phonon family. Ollama also lists French, German, Spanish, Portuguese and multilingual variants. (Phonon on Ollama)
However, the models do not all carry the same licence:
| Model group | Published licence |
|---|---|
| Phonon 31M English | Apache 2.0 |
| Phonon 40M language-specific models | CC BY-NC 4.0 |
| Phonon 40M multilingual | CC BY-NC 4.0 |
Gradium’s model cards identify these differences, including the noncommercial restriction on the French 40M checkpoint. Developers considering a commercial product should check the licence of the exact model they intend to distribute or use. (English model card, French model card)
Language availability and permission for a particular use are separate checks.
Local speech does not make the whole assistant local
Gradium presents Phonon as an on-device speech system: generation can happen without sending the text to a remote speech service. That can support offline applications and reduce dependence on a network connection for the speaking stage. (Gradium’s Phonon release)
The rest of the application still matters.
If a cloud language model writes the answer, that part of the workflow remains remote. If an application uploads documents, records conversations or synchronizes logs, adding local speech generation does not change those behaviours.
For privacy-sensitive projects, examine the complete route from input document to spoken response.
How to interpret the performance claims
Ollama reports that its implementation was 35% faster than Gradium’s runtime in its comparison on a MacBook Pro with M5 Max.
The published conditions describe generating 45 seconds of speech and taking the median of ten runs. The implementations also used different numerical formats. This is a vendor-published result for a specific setup, rather than a performance guarantee for every Mac. (Ollama’s benchmark conditions)
For a practical test, separate three questions: how quickly speech starts, whether generation keeps up with playback, and whether the spoken result is accurate and comfortable to hear.
Check difficult vocabulary and longer passages as well as speed.
A useful addition to the local AI toolbox
Our assessment is that Phonon’s integration matters because it makes speech generation easier to include in an Ollama-based application.
The strongest starting points are focused: reading an answer aloud, testing an interface or producing a short narration sample.
Before building further, verify the supported hardware, select the right language checkpoint and review its licence. Then evaluate the model with the text your application will actually speak.