AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Speech models in Ollama

Collected Oct 25, 2026

Ollama has added text-to-speech to its local model runner, with Phonon from Gradium as the first supported speech model. The feature is available on macOS now, with Windows and Linux promised later. Users need Ollama 0.40.11 or later and pull the model with "ollama pull phonon"; the simplest invocation is "ollama run phonon" followed by the text to speak.

Phonon is described as a family of small, state-of-the-art text-to-speech models. It supports five languages and dozens of voices drawn from Gradium's voice catalog, and output streams as it is generated. On Apple Silicon, Ollama's implementation is accelerated by MLX, Apple's machine learning framework. On a MacBook Pro with an M5 Max, Ollama measured itself at 35% faster than the next fastest engine, Gradium's own runtime, and says its accuracy matches the reference implementation in Hugging Face Transformers as measured by word error rate. For context, word error rate is the standard metric for transcription and speech quality: the share of words a system gets wrong, so a match here means Ollama's faster path did not trade accuracy for speed. The headline claim of more than 100 times faster than real time means a minute of audio can be synthesized in well under a second on that hardware, though the source ties that figure to Apple Silicon generally rather than to one machine.

Voice selection works through a --voice flag on the command line or a voice field in the API. Running "ollama show phonon" lists every available voice. Licensing is split: phonon itself is Apache 2.0, but the 40M models carry CC BY-NC 4.0, which restricts them to noncommercial use. That distinction matters for anyone planning to ship a product, since the permissive code license does not extend to the downloadable weights.

A practical pattern is chaining text and speech models. Piping a chat model's reply into Phonon reads it aloud, and because text is spoken as it arrives, playback can begin after the first sentence rather than waiting for the full response. Ollama gives the example of piping output from gemma4 into phonon. Running phonon with no text drops into a prompt for typing lines interactively, and redirecting output saves a WAV file.

The API surface is deliberately familiar. Phonon works with the OpenAI-compatible /v1/audio/speech endpoint, so existing code using the OpenAI SDK needs only a new base URL pointing at the local server. Ollama shows curl, Python and Node.js examples: in Python, the streaming response create call writes straight to a file; in Node, the audio buffer is written with the filesystem module. Responses stream while audio generates, so a player like ffplay can start immediately as bytes arrive. Audio is 24 kHz mono, returned as WAV by default or as raw 16-bit samples when the response format is set to pcm.

This suggests Ollama is positioning speech as another local primitive alongside chat models, and the OpenAI-compatible endpoint lowers the cost of switching: teams already calling a hosted speech API can redirect to localhost and keep their code, gaining offline operation and per-call cost removal at the expense of running inference on their own hardware. A likely trade-off is hardware dependence: MLX acceleration ties the fast path to Apple Silicon Macs for now, so the performance numbers do not transfer to Windows or Linux machines until those builds arrive. A second trade-off is voice breadth and licensing: the noncommercial terms on the 40M models mean commercial products may need different arrangements with Gradium. The pipeline behavior, where speech starts before the text model finishes, is the most developer-relevant detail, since it makes voice interfaces feel responsive without waiting for a complete reply.

Why it matters: developers building assistants, accessibility features or narration tools can now add speech synthesis locally through an interface they may already use, with no per-character charges. Product teams evaluating voice output should check hardware support and the CC BY-NC 4.0 terms on the model weights before committing to Phonon; the Apache 2.0 code license alone does not cover commercial use of those models. Those on Intel Macs or other platforms will likely have to wait for the upcoming Windows and Linux releases, which is an inference from the stated platform support rather than a confirmed timeline.

Read at Ollama

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Ollama now supports speech models, starting with Phonon from Gradium. Read text aloud in five languages and dozens of voices, streamed as it is generated, over 100 times faster than real time on Apple Silicon.