Use Your NiceVois Training Voice

Your voice is saved in your library. Convert with it on NiceVois, or keep its complete ZIP for a matching Seed-VC setup.

Convert with your voiceChoose your saved voice, add a vocal or song, set the pitch, and press Convert.

Keep the complete ZIP

It contains your trained voice, its settings, a short reference recording, a license, and file checksums. Keep them together. There is no RVC index. The model.pth inside this ZIP is a Seed-VC checkpoint, not an RVC model. Do not load it in Applio or an RVC-only interface.

Use it on NiceVois

Choose your voice in Voice Conversion. You can use the same pitch and output-format controls as Standard RVC. Keep pitch at 0 unless you want to change the key. If you move the voice to another NiceVois account you control, upload the complete voice ZIP, not its individual checkpoint.

Use it on your computer

This voice uses Seed-VC V1 singing conversion. A compatible GPU setup is needed for practical speed. Follow the official installation instructions for that revision, then unzip your voice into a private folder.

python app_svc.py --checkpoint /path/to/voice/model.pth --config /path/to/voice/config.yml --fp16 True

Open http://localhost:7860/. Use your vocal as Source Audio and the included reference.wav as Reference Audio. Start with 30 diffusion steps, length 1, guidance 0.7, automatic pitch adjustment off, and pitch shift 0.

You can also convert from the command line. We tested a downloaded NiceVois voice ZIP in the matching upstream code on a separate Linux GPU runtime:

python inference.py --source /path/to/vocal.wav --target /path/to/voice/reference.wav --output /path/to/results --checkpoint /path/to/voice/model.pth --config /path/to/voice/config.yml --f0-condition True --auto-f0-adjust False --semi-tone-shift 0 --diffusion-steps 30 --length-adjust 1 --inference-cfg-rate 0.7 --fp16 True

The web interface needs Gradio 5.23.0, as listed in the matching upstream requirements. Installation and speed can vary by computer; the cloud version does not require you to install anything.

The local sound can differ. NiceVois also applies the cleaner audio finish selected in our tests. Its weight is not bundled in your voice ZIP. The upstream interface returns Seed-VC audio without that NiceVois finishing pass.

Preparing another voice

Start with 2–10 minutes of clear solo voice. Avoid music, heavy echo, effects, and overlapping singers. You can add one recording or several separate files. NiceVois chooses the settings; you do not need to choose epochs. One new voice uses one training pass or one subscription training, just like Standard RVC. The trainer shows your free training left for each method.

NiceVois Training or Standard RVC?

Both create a saved voice from your recordings. NiceVois Training uses Seed-VC and chooses the settings for you. In our listening tests it preserved sung words more clearly. Results still depend on your recordings and the vocal you convert.

Standard RVC keeps the familiar epoch controls and .pth/.index files for RVC tools. NiceVois Training gives you a different complete ZIP for Seed-VC. On NiceVois, both use the same voice picker, pitch control, Convert button, and plan terms.

Credits and source

Built on Seed-VC, with NiceVois training, validation, and audio-finishing changes. Seed-VC publishes GPL-3.0 terms. The finishing code is from PlayVoice BigVGAN; its separate weight terms are awaiting clarification. No license grant for that weight is claimed.

Download the matching engine source and license notices. This source archive contains no voice recordings, trained customer voices, or finishing weights.

Only use recordings and voices you have permission to process. Keep the reference recording and model private unless you have permission to share them.