Skip to content
Roboto
Log in Try Roboto

OmniVoice: clone your voice locally in French, key settings and licence limits

Cloning your own voice without sending recordings to a server is what OmniVoice promises. The model, developed by Xiaomi's Next-gen Kaldi team, needs only 3 to 10 seconds of reference audio to produce a synthetic voice that sounds like yours, and it runs locally, for free, on Mac, Linux or Windows. Two important caveats before you start: output quality depends heavily on how clean your reference clip is, and the model weights are under a CC-BY-NC licence, which rules out commercial use. Here is what you need to know to install it, configure it properly and avoid the common pitfalls.

What OmniVoice does, and what it implies

The model supports 646 languages according to the source, and French is not an afterthought: it is the fifth language in the training corpus, with 23,675 hours out of a total of 581,000. In practice, that means French pronunciation is better covered than in many multilingual tools, though it does not guarantee perfect results on proper nouns or complex sentences.

Usage is local and free, but two licences are at play. OmniVoice's code is released under Apache 2.0, a permissive licence. The model weights, however, are under CC-BY-NC: personal use allowed, commercial use forbidden. This is a nuance the README reportedly does not highlight, and it changes everything for a creator hoping to monetise generated voiceovers.

Installation: prerequisites to check first

OmniVoice requires Python 3.10 or newer, inside an isolated environment. The first run downloads roughly 3.3 GB of weights, so plan for disk space and a decent connection.

Hardware matters more. On Mac, the model uses the integrated GPU via MPS. On Linux and Windows, an NVIDIA card is needed for comfortable use. Without a GPU, the installation still works, but computation takes about five times the audio duration according to a test cited by the source. In other words, a ten-second sentence can take close to a minute.

Be careful with the project's performance claims too: the 40x real-time factor was measured on an H100, not on a desktop machine. On an M4 Max, the source counted 4 to 6 seconds of computation for 5.4 seconds of audio. That is usable, but far from the headline figure.

The two settings that really change the result

1. The quality of the reference clip

This is the most decisive factor. The clip must be clean: no silence, no hesitation, no background music. Beyond 10 seconds, the clip slows down computation and degrades cloning according to the source. A short, sharp passage beats a long, sloppy recording.

2. Providing the exact transcription

If you do not supply the transcription of the clip, the tool downloads Whisper, an extra 1.6 GB, and may get proper nouns wrong. Providing the exact text avoids both the download and those errors. One practical detail: the command line applies no normalisation, so numbers must be written out in full.

Limits to know before committing

Cloning is not flawless. Out of 24 generations of the same sentence, 5 lost or distorted a word, often at the beginning or the end, according to the source. This is a known bug that forces you to retry several times to get a clean take. Quality also depends heavily on how clean the reference audio clip is.

Above all, the CC-BY-NC licence on the weights closes the door to professional projects. For personal, experimental or educational use, OmniVoice remains an interesting option. For commercial voiceover production, you will need to look elsewhere or check the terms with the team.

In short

  • What it is: a local, free, multilingual voice cloning model (646 languages), developed by Next-gen Kaldi (Xiaomi).
  • What you need: Python 3.10+, an isolated environment, 3.3 GB of space, and ideally an NVIDIA card (or a recent Mac).
  • The two key settings: a clean 3-to-10-second reference clip, and the exact transcription supplied manually.
  • The limit: weights under CC-BY-NC, so no commercial use.
  • Real-world performance: far from the advertised 40x real-time, which was measured on an H100.

OmniVoice is a promising tool for anyone who wants to experiment with voice cloning in French without relying on a cloud service. Provided you accept its limits: occasional generation glitches, quality that hinges on the reference clip, and a licence that seriously restricts professional use.

Source

OmniVoice - Clonez votre propre voix en local et en français