OpenVoice
nigiri.openvoice
A Docker Compose file for OpenVoice alone, generated from the catalog and checked by the safety inspector. No account needed.
Voice cloning system that reproduces a reference speaker's tone from a short sample and applies it to synthesised speech, including in languages absent from its training data.
OpenVoice separates what is said from who appears to be saying it: a base speaker model generates speech with the intended language and style, and a tone colour converter then transfers the timbre of a reference recording onto that output. Because the two stages are independent, a voice captured in one language can speak in another the model never heard it produce, and style parameters such as emotion, rhythm, pausing and intonation stay controllable after cloning. Version 2 improved audio quality and added native support for six languages. A few seconds of clean reference audio are enough for a usable clone.
You know it worked when
- The service starts and reports its checkpoints loaded.
- A synthesis request with the base speaker returns intelligible audio.
- A reference sample produces output recognisably in that speaker's timbre.
- The same cloned voice speaks a second supported language.
- A style parameter such as emotion produces an audible difference in the output.
Known sharp edges
- Cloning a real person's voice carries consent, likeness and impersonation exposure that no software licence addresses, and several jurisdictions now regulate it directly.
- Output quality depends heavily on the reference recording: background noise, compression artefacts or a short sample produce an audibly wrong clone rather than a failure.
- Accent and dialect are not reliably preserved — a distinctive regional voice frequently comes back flattened toward the base speaker's accent.
- Model checkpoints are downloaded separately from the code and are not included in the image, so a fresh container generates nothing until they are placed.