Check a solo voice recording before turning it into a portable RVC model. You can see how much audio is active, how much is silence and whether the samples hit full scale.
What to record for a singing voice model
Use clean recordings of one voice you own or have permission to use. Varied solo singing across comfortable low and high notes is more useful than repeating one note. Two to five minutes is a practical starting point. A short clean clip can still train, but it may not cover the notes you want to sing.
What the measurements mean
- Active audio: an estimate based on signal level. It can include breaths, background sound or instruments; it is not a verified count of singing seconds.
- Silent gaps: quiet intervals the existing NiceVois silence detector would remove. Listen before trimming; quiet singing may be intentional.
- Clipped samples: samples very near full scale. A high share can indicate a recording made too loud; peak level alone does not prove distortion.
- Vocal pitch: a pitch estimate on voiced frames, useful for checking a solo recording's register. Unpitched consonants and silence are excluded.
Prepare the recording, then train
If you already know the file contains music, use the vocal remover. For layered vocals, use the lead and backing splitter. Hiss and echo have different tools: noise removal and reverb removal. This checker does not decide whether a file contains music or harmonies.
When your own recording is ready, train a voice or follow the dataset preparation guide. The existing free training is one run per signed-in account, up to five minutes of audio and 150 epochs. Download free-run models within their seven-day storage period.
Common questions
Is this free, and is my recording uploaded?
The check is free and runs locally in your browser. No account is needed. The recording is not uploaded for analysis. Anonymous usage events record the check's outcome and broad measurements, not your audio or file name.
Will it guarantee a good voice clone?
No. The measurements can reveal some recording issues, but they do not judge how natural a trained voice will sound. Listen to the source, use one permitted voice and vary the material.
What if my file will not open?
Audio formats supported by your browser vary. Try a WAV or MP3, or use the audio format converter. The checker accepts files up to 100 MB and ten minutes; split longer material before checking.