MODEL COMPARISON

OmniVoice vs SuperClonica: more languages or more speed

OmniVoice, from the k2-fsa project, is a zero-shot text-to-speech model with very wide language coverage: its documentation lists 646 languages. It clones a voice from a few seconds of audio and can design one from a list of attributes. SuperClonica covers far fewer languages, 31, but runs fast on a regular CPU and keeps a voice stable from line to line.

Here is how the two compare and when each is the right choice.

Hear voices that run on CPU Set up the SuperClonica engine

What OmniVoice does well

OmniVoice uses an architecture in the style of a diffusion language model, with Qwen3-0.6B as its base model.

  • Languages: 646 in the official list, with 581,000 hours of training data.
  • Voice cloning from a 3–10 second reference clip and its transcript, without a training step.
  • Voice design from attributes such as gender, age, pitch, accent or whispering. The authors note that this mode was trained on Chinese and English data only and may be unstable in some low-resource languages.
  • GPU speed: the authors report a real-time factor as low as 0.025, which is 40× faster than real time.
  • Hardware: NVIDIA GPUs, Apple Silicon and Intel Arc are supported.

OmniVoice on CPU: our speed estimate

On a desktop 8-core CPU we estimate OmniVoice at 2–5× slower than real time, so a 10-second phrase takes 20–50 seconds. SuperClonica does it in about 1.0–1.6 seconds, which is 6.4–9.7× faster than real time.

The estimate starts from omnivoice-server, an HTTP server for OmniVoice, which reports a real-time factor of 4.92 on CPU at 32 steps without naming the processor. We assume up to 2.5× more speed on a fast desktop CPU. Like every speed in our table, this is an estimate, not a benchmark.

Quality and timbre stability

Our table gives OmniVoice and SuperClonica the same quality mark, 4.5 out of 5. The difference lies elsewhere: CPU performance (2 against 5) and timbre stability (4 against 5). A stable timbre matters when one character reads hundreds of lines, as in games and audiobooks.

OmniVoice's own documentation points the same way: cloning is its most stable mode, and longer reference clips slow it down and may hurt the result.

License

The OmniVoice code is released under Apache 2.0, but its model card says the pre-trained model is licensed CC-BY-NC because of the training data. Under the default terms, that rules out commercial use of the official weights.

Supertonic 3 uses OpenRAIL-M, a royalty-free license with a list of prohibited uses instead of a non-commercial clause. Voices bought on SuperClonica are for your own projects and may not be resold.

OmniVoice or SuperClonica for text-to-speech?

Decide by language first, then by hardware.

  • OmniVoice: your language is not among the 31 that Supertonic 3 supports, you need Chinese or instant cloning, and you have a GPU.
  • SuperClonica: speech has to run on CPUs or inside a browser or app, offline and quickly, with the same voice across thousands of lines.