A safe place to start
These are common starting points, not magic numbers. Transpose: 0 for a similar voice, plus 12 for a male vocal into a female model, minus 12 for the reverse. Pitch extraction: RMVPE. Index rate: around 0.5 to 0.75. Protect: 0.33. Filter radius: 3. RMS mix rate: around 0.25.
Convert one short phrase with these values, listen, then change one setting at a time. Changing three at once makes it impossible to know which one helped.
Transpose
Transpose shifts the pitch in semitones before conversion, so the input sits in the model's natural range. Twelve semitones is one octave. A male vocal pushed through a female model usually needs plus 12; a female vocal into a male model usually needs minus 12. Smaller shifts, like plus or minus 3 to 5, suit voices that are close but not matched.
If the result sounds strained or cartoonish, the transpose is probably wrong, not the model.
Pitch extraction method
Before converting, RVC has to track the melody of the input. RMVPE is the usual choice today: fast and reliable on most vocals. Older options such as Harvest and Crepe still appear in many tools and can help in specific cases, like very breathy or unusual voices, but they are slower. If notes wobble or jump an octave, trying a different method is a quick test.
Index rate
RVC starts from a pretrained base model, and a little of that base, or of the input voice, can leak into the result. The index rate controls how strongly the model pulls toward the features of its own training audio. Higher values sound more like the trained voice. Lower values can sound cleaner when that training audio was noisy, because the noise lives in the index too.
If a conversion sounds breathy, gritty, or full of small artifacts, lower the index rate first. If it does not sound enough like the target voice, raise it.
Protect, filter radius, and RMS mix
Protect keeps breaths and unvoiced consonants, like s, t, and k, from being converted into strange tones. Lower values protect more; 0.5 effectively turns it off. Too low and words can start to sound swallowed, so 0.33 is a sensible middle.
Filter radius smooths the tracked pitch to reduce breathiness and wobble; 3 is the usual default. RMS mix rate decides whether the output follows the input's loudness or the training voice's. Lower values keep your original performance's dynamics, which is usually what you want in a song.
Speech vs singing
For songs, keep transpose to whole octaves: 0, plus 12, or minus 12. Any other shift moves the vocal into a different key, and it will clash with the beat. If you need a different key, change the whole song afterwards with a pitch changer instead.
For speech, small transposes are fine, because there is no beat to match. Spoken lines also benefit from a slightly lower protect value, so consonants stay crisp, and a moderate index rate, since speech exposes artifacts more than a busy mix does. Always test with the kind of material you will actually use: a sung test line says little about how a voiceover will sound.
Fix it by symptom
Metallic or robotic: the input probably has reverb, backing vocals, or music in it. Separate the vocal with the stem splitter and remove hiss with the noise reducer, then try again. Wrong octave or strained: adjust transpose. Mushy s and t sounds: lower protect slightly. Volume pumping: lower the RMS mix rate. Artifacts on every word: lower the index rate.
If nothing helps, the model itself may be undertrained or overtrained; our guide on how to train an RVC model covers what to change.
The setting nobody lists: the input
Every setting above works on top of the vocal you feed in. A dry, close-miked, in-tune vocal will convert cleanly with defaults. A vocal with room echo, doubled harmonies, or a beat underneath will fight you no matter what you change.
In Xport Studio the vocal converter exposes these same controls with sensible defaults, and the stem splitter and noise removal sit in the same app, so cleaning the input takes one step instead of three tools.