Ollama adds local speech, with a license difference between models

Creators who already run models on a Mac have another way to draft narration. Ollama’s October 24 speech release adds Gradium’s Phonon text-to-speech family for Apple Silicon Macs, with audio streamed as it is generated. Launch announcement.

Before choosing a language or model size, check the exact checkpoint. The model card lists the default 31M English model under Apache 2.0, while the 40M checkpoints use CC BY-NC 4.0, which restricts them to noncommercial use. A different tag is therefore more than a voice-quality preference when planning paid work.

Test the sentence your audience must understand

For an illustrative first draft, use a short paragraph containing a product name, a price and a written-out date. Compare what is spoken with the approved text. A pleasant voice is not enough if it changes the name or makes the price ambiguous. CreatorStack has not generated or listened to this example.

If the result is difficult to follow, revise the sentence and punctuation before committing to a longer script. Keep one approved sentence as a reference so you can judge whether a later change actually helps. For final publication, listen again in the surrounding edit rather than approving the audio in isolation.

The speech documentation describes WAV and PCM output and a 4,096-character limit per API request. It also says unsupported settings fail rather than silently taking effect. Do not assume controls from another speech service work unchanged.

Match the setup to this release

The launch recommends Ollama 0.40.11 or later and describes Windows/Linux support as forthcoming. Other current documentation lists a lower minimum version and cross-platform playback helpers; those details do not establish that this launch’s model packages run on every platform. Use the newer launch recommendation on a supported Mac and confirm compatibility before building a workflow around it.

This is speech generation, not transcription or a complete voice assistant. Start with the intended output and model license, then judge the reading against your script. Our AI output evaluation guide explains that review; the ElevenLabs guide covers a hosted audio workflow for comparison.