Skip to main content

Engines

An engine is a Python package that the server installs into a virtualenv of its own, through the API, the web page or pnpm wizard install. One engine's dependencies cannot break another's, and a crash takes down that engine's worker rather than the server.

Which one​

  • Start with Kokoro. It runs on the CPU with no torch, its weights are about 200 MB, and it speaks faster than realtime on a laptop. It performs no cues, and makes new voices by blending its own rather than by cloning.
  • For the fastest speech on a CPU, Phonon. It streams at more than 50 times realtime from a 75 MB download, with no torch, in 38 voices. It is English only, and performs no cues. Phonon 40M is the same runtime in French, German, Spanish and Portuguese too, for non-commercial use.
  • With a GPU, Chatterbox. It clones a voice from a few seconds of audio. Its turbo and nano builds perform all eight cues; its original and multilingual builds take dials instead. Which you get depends on the build that is loaded, and the capability document says which.
  • For audio that starts before the line is finished, Orpheus. It streams while it generates, and performs seven of the eight cues. On Linux it runs the full build through vLLM on an NVIDIA card; elsewhere the q8 and q4 builds run through llama.cpp.
  • For two people talking, Dia. It speaks a two-speaker dialogue in one take, performs all eight cues, and clones from a clip and the words spoken in it. It wants a real GPU: on a Mac it runs at about a tenth of realtime.
  • For the most natural reading, if you are not selling it, Breeze TTS 2. It ranked first among open weights on the Artificial Analysis speech arena at its release, speaks English and Chinese, performs all eight cues, and clones from a clip and its transcript. Its weights are for research and non-commercial use only, and it needs an NVIDIA card with about 8 GB free.
  • For more languages, if you are not selling it, Fish Audio S2 Pro. Built with Fish Audio, it ranked second among open weights on the Artificial Analysis speech arena at its release. It speaks ten languages, performs three of the eight cues, and clones from a clip and its transcript. Its code and its weights are for research and non-commercial use only, and it needs a card with about 13 GB free.
  • For a clone from nothing but a short clip, without a GPU, StyleTTS 2. It clones from three seconds of audio with no transcript, runs several times faster than realtime on a laptop's CPU, and holds 1.2 GiB. It speaks English only and performs no cues. Its weights are 0.9 GB, and upstream asks that listeners be told the speech is synthesised unless the person cloned has agreed.
  • Tone makes a sine wave. It exists to prove the protocol in CI, and it is the fixture the conformance suite compares a real engine against.

Licences​

Every engine carries two licences, one for its code and one for its weights, and the catalog shows both before anything is installed. An engine whose weights may not be used commercially installs only once that licence is accepted by name. The core is MIT; each engine is a separate package under its own licences, so an engine with restrictive weights can sit in the catalog without touching the core.

Adding one​

An engine is one Python class. See adding an engine.