Ollama added text-to-speech on October 24, bringing speech generation to Apple Silicon Macs through Gradium’s Phonon model family. Phonon is the first family of speech models available through Ollama, and users need version 0.40.11 or later to run it. Ollama says support for Windows and Linux is coming, without specifying when.
The licensing differs between models: phonon carries an Apache 2.0 license, while the 40M models use CC BY-NC 4.0 and prohibit commercial use. The standard phonon model produces English audio. For speech in other languages, separate 40M models cover Portuguese, German, French and Spanish. The phonon:40m-multilingual option handles all five supported languages in one model.
Users can download the default model by running ollama pull phonon. Running ollama run phonon without supplying text lets users enter lines for speech generation. Sending the command’s output to a file saves the generated speech in WAV format.
Users can also pass a chat model’s replies straight to Phonon for spoken output. In that setup, playback starts once the reply’s first sentence is ready. For applications using the API, audio streams as it is generated so playback can begin before generation finishes.
Developers can access Phonon through the OpenAI-compatible /v1/audio/speech endpoint. Ollama says existing code using the OpenAI SDK needs only a different base URL to work with it. The announcement includes API examples in Python, JavaScript and cURL.
Voice choices come from Gradium’s catalog, and the command ollama show phonon lists the available options. Applications can choose a voice through the API’s voice field. Hadley and Zoey are among the voices demonstrated in Ollama’s examples.
The API returns WAV audio by default; setting response_format to pcm produces raw audio samples instead. Acceleration on Apple Silicon relies on Apple’s MLX machine learning framework. Ollama reports a word error rate equal to that of the Hugging Face Transformers reference implementation.
The speech addition follows Ollama’s October 13 introduction of cloud model access inside Fly.io computers. An earlier update on September 29 brought decision models using TypeSafe’s Jev API, including models for yes-or-no responses, scoring and choosing between options.
