No API key and no cloud call: the model is fetched once and everything after that is local, which makes it usable on text you would not hand to a hosted service. Voices and expression tags are enumerable, so the assistant can pick one rather than being told which to use.
A local text-to-speech server powered by Supertonic 3, running over stdio. Synthesis happens on your own machine; the model downloads once into a cache and works offline after that.
- Text synthesised to a WAV file, returned as an absolute path with its metadata
- Built-in voices listed with their IDs and gender
- Inline expression tags - laugh, breath and others - listed with descriptions, usable inside the text
- A language set per call, so non-English text is spoken correctly rather than read as English
- Optional playback on the machine running the server, on macOS and Linux
- Offline operation after a preload step that downloads the weights and prints their checksums
Python and a virtual environment, installed from a checkout. Playback and synthesis output paths cover macOS and Linux; Windows is not supported in the current version. The first start downloads the model unless you preload it.
Build from source — clone the repository and build it, then point your client at the binary