Ollama 0.40.9 Adds Streaming Speech Synthesis; Phonon Initially Limited to Apple Silicon

Ollama’s October 22 release, version 0.40.9, adds Phonon text-to-speech support, allowing local applications to receive audio as it is generated through the command line or the /v1/audio/speech endpoint. This adds a speech output path to existing chat applications. The release also switches model imports to the production Gradium checkpoint format, so compatibility with the older preview format may require attention. Release notes

The default phonon model is an English model with 31 million parameters and a voice named default. The official package currently requires Apple Silicon and Ollama 0.40.9 or later. There are also 40-million-parameter English, French, German, Spanish, and Portuguese models, as well as a multilingual version covering five languages that do not include Chinese. Check the licenses for each model: the default English model uses Apache 2.0, while the 40-million-parameter series uses CC BY-NC 4.0, which restricts commercial use. Model details

For integration, developers can connect a chat model’s text output pipeline to ollama run phonon or use a compatible SDK to call the speech endpoint. Output supports 24 kHz, mono, 16-bit WAV or raw PCM. The API accepts up to 4,096 characters per request, while the command line splits long text by paragraph and sentence. WAV streams send a header before the total duration is known, using the maximum value in the length field. Receiving applications should confirm that their players and file-handling workflows can accept this format. Speech documentation

From an interface design perspective, this update can reduce the integration work needed to connect speech services to local chat systems, but a complete experience still depends on text segmentation, audio buffering, and playback scheduling. The official documentation also specifies that numbers, dates, and abbreviations are not automatically expanded into spoken forms; unsupported speed, format, or instruction options return a 400 error. Engineering teams should test first-chunk audio latency, continuation with long text, and playback cancellation, and treat Chinese support and other hardware platforms as items to confirm. Public information is currently insufficient to draw cross-hardware performance conclusions. Interface limitations