Free & private · nothing is uploaded

RVC Dataset Maker

Turn long vocal recordings into RVC training clips. It cuts at the silences, shows every cut, and gives you a ZIP of WAV files.

0clips
0:00clean minutes
0:00left out

    How to make an RVC dataset from long recordings

    1. Drop in one or more vocal recordings (WAV, MP3, M4A, or FLAC). Clean, dry vocals work best.
    2. Check the waveform. Each colored block is one clip, and the gray parts are silence that gets left out.
    3. Adjust the minimum and maximum clip length or the silence threshold if the cuts look wrong. The clips update as you type.
    4. Play a few clips, then press Download ZIP. You get clip_0001.wav and up, mono 16-bit, ready for an RVC trainer.

    How much audio RVC needs

    The RVC project's own FAQ recommends around 10 to 50 minutes of audio. It says 5 to 10 minutes can be enough for a distinctive voice on a clean recording, and that results from under a minute have not been reproducible. The clean minutes counter above adds up only what lands in the clips, so you can see where your dataset stands before you train. Our guide on how to train an RVC model covers the rest of the process, and the best RVC settings explains epochs, pitch, and index rate.

    Length only helps when the audio stays clean: one singer, one tone, no music, no reverb. Ten minutes of dry vocals usually beat an hour of live recordings. If your vocal is inside a finished song, pull it out with the stem splitter first, and take out hiss or hum with the noise reducer.

    How the slicing works

    The tool measures the loudness of the audio every 10 milliseconds. Wherever it stays below the silence threshold for at least 300 milliseconds, that pause becomes a cut, placed on the quietest moment and keeping at most half a second of silence on each side. A piece that runs past the maximum length is cut again at its quietest point, and anything shorter than the minimum is left out. This is the method of slicer2, the open-source audio slicer from OpenVPI that RVC's own preprocessing is built on, rewritten to run in your browser.

    Private by design

    Your recordings are decoded, sliced, and zipped on your own device. Nothing is uploaded, which matters when the dataset is your own voice. Files the browser cannot read on its own are decoded with FFmpeg compiled to WebAssembly, which also runs locally and downloads once, about 31 MB.

    RVC dataset maker: common questions

    How many minutes of audio does RVC need?

    The RVC project's FAQ recommends around 10 to 50 minutes of audio. It says 5 to 10 minutes can be enough when the voice is distinctive and the recording is clean, and that results from under a minute have not been reproducible. Clean audio matters more than length.

    Why slice at the silences?

    A cut in the middle of a word leaves a clipped sound at the start or end of a clip. Cutting where the singer breathes or pauses keeps every clip a whole phrase, and the silence between phrases is left out of the dataset.

    What clip length should I use?

    The default of 3 to 12 seconds keeps each clip one or two phrases, short enough to check by ear and delete if it is bad. RVC's own preprocessing cuts training audio into shorter pieces again, so the exact length is not critical.

    What silence threshold should I pick?

    Start at -40 dB. If long stretches come out as one piece because the background never gets quiet, raise it toward -30 dB. If quiet words get cut off, lower it toward -50 dB. The clips and waveform update as you change it.

    What happens to a long note with no pause?

    If a stretch runs past the maximum length without a silence, the tool cuts it at the quietest point it can find inside the allowed range, so no clip is longer than the maximum.

    Should I use 40 kHz or the original sample rate?

    40 kHz mono matches RVC's 40k model setting, and training resamples the audio to the model's rate either way. Keep the original rate if you plan to train a 48k model from a 48 kHz recording.

    Should I clean the vocals first?

    Yes. The model learns whatever is in the clips, including music, reverb, and noise. Pull the vocal out of a song with the stem splitter, reduce hiss or hum with the noise reducer, and only then slice it here.

    Can I train in Applio, the RVC WebUI, or Xport Studio with these clips?

    Yes. The ZIP holds plain 16-bit mono WAV files named clip_0001.wav and up, which any RVC trainer reads as a dataset folder. Xport Studio's Train Clone takes up to 100 files per voice.

    Is my audio uploaded?

    No. Your browser decodes the files, finds the silences, and builds the ZIP on your own device. Nothing is sent to us or anyone else.

    Who makes this?

    We're the small team behind Xport Studio, a private desktop app for AI vocals, stems, and mastering that processes your audio on your own machine. These free tools run on the same idea: your audio stays on your device. If you're curious about the bigger thing, take a look.

    Meet Xport Studio