Vocal Isolation — Remove Backing Music and Room Noise from a Vocal (No Upload)

00:00.00
00:00.00

Free online vocal isolation — extract vocals and reduce noise in your browser, no upload. Record or upload in the header Controller, pick a mode, press Isolate vocal, compare on the pitch roadmap, then export with Download Isolated (MP3, WAV, OGG, or M4A) or Save to take into slot A–D.

Your voice never leaves your device

Separation and noise reduction run locally with WebAssembly and ONNX Runtime Web. The audio stays in your browser tab and is never sent to a server. A slow connection does not add queue time — only the one-time model download (about 30 MB) uses the network. Longer takes take longer to process, and the progress bar shows time left.

Hear an acapella-style vocal without a DAW

This won't replace iZotope RX or a paid stem service on a dense master. Separate vocal from mix estimates the lead vocal over drums, bass, or chords. Reduce room noise cleans steady hiss, hum, or fan noise on a dry recording — it does not remove music. Use it to practice pitch on a song fragment, check intonation on a noisy home recording, or preview what an isolated vocal might sound like before you invest in studio tools.

Play the original and the isolated version back to back

After Isolate vocal, use Load Original and Load Isolated to switch versions in the header player at the same timestamp. The isolated file stays on this page until you press Save to take, choose a slot, and name it. The pitch roadmap shows the original trace in gray and the isolated trace in teal. Click the chart to move the playhead.

Download the isolated file without leaving the browser

After processing, use Download Isolated beside Re-isolate: pick Format (MP3, WAV, OGG, M4A), then download. Filenames use the take’s recording time, e.g. isolated_2025-10-03_14-30-00.wav. Or press Save to take, pick slot A–D, and export that slot from the Controller download menu — same local encode, no upload.

Chain with Auto-Tune when you need both

Each tool processes the take you load in the Controller, not a hidden copy tied to another slot. Fastest path: Save to take after isolation, select that slot, open Auto-Tune, and Apply. You can also download the isolated file (WAV keeps the most detail), upload it as a new take, and run the second tool — same idea as chaining files in a DAW, but entirely in the tab.

Mode

Separation estimates the lead vocal from a mixed song. Drums or reverb can still leak in on dense mixes. Pick the range of the singer in this file. It may not be your own range.

Enable the microphone and record a take, or upload an audio file. The file becomes a take in the header Controller, then Isolate vocal works on it.

Example of the panels that appear after processing. No take is loaded.

Separate vocal from mix · Length 3.2 s · Vocal level -4.1 dB · Removed energy -2.6 dB

Detected range: A3–D4

good for pitch practice

Range–

Top = pitch overlay; bottom = correction (cents, magnified).

  • Original
  • Isolated
  • Pink band / ticks = gap between curves
  • Gold line = playhead (matches pitch monitor)

How to use it

Record a take in the Controller strip in the header, or upload a file (up to 10 minutes / 50 MB). Choose Separate vocal from mix for a song with backing music, or Reduce room noise for a voice-only take. Optionally pick the vocal range, then press the button. The time estimate shows before you start; the progress bar shows the real work done.

Vocal range (optional)

  • Pick the range of the singer in this file. It may not be your own range — for a commercial song it is the original singer’s.
  • The range does not make processing faster. It narrows the pitch trace to those notes (±2 semitones) and removes rumble below the lowest note, which cuts false pitches from bass left in the mix.
  • Auto estimates the range from the result and shows it as Detected range. If too few sung notes are found, no range filter is applied.
  • Change the range after processing and press Apply new range: only the range filter and the pitch trace run again, not the model. After a reload, Process again is needed.
  • If most of the vocal sits outside the chosen range, the result card warns and offers Switch to Auto.

After processing

  • Pitch roadmap — gray is the original pitch, teal is the pitch measured on the isolated file. Click the chart to move the playhead. Gaps are unvoiced sections (consonants, breath).
  • Load Original / Load Isolated — pushes either version into the Controller player so you can compare them at the same timestamp.
  • Re-isolate — change mode, vocal range, or room-noise sample and run again. Always starts from the original audio; passes do not stack.
  • Download Isolated — appears in the action row after Isolate vocal: choose MP3, WAV, OGG, or M4A. Saves locally; filename uses the take recording timestamp (isolated_YYYY-MM-DD_HH-MM-SS.ext).
  • Save to take — stores the isolated audio in a slot A–D with a name you choose. It does not replace the dry or mixed recording in another slot unless you overwrite that slot on purpose.
  • Controller export — after Save to take, open Download in the header, pick that slot, and choose MP3, WAV, OGG, or M4A.

Reading the result card

  • Vocal level — loudness of the isolated file against the original, in dB.
  • Removed energy — loudness of what was taken away (original minus isolated), in dB.
  • Noise floor change — in Reduce room noise, how much quieter the noise sample got.
  • Good for pitch practice means the gaps between phrases are quiet. Backing still audible means instruments remain between phrases; a voice-only take works better. A mono mix (both channels the same) separates less cleanly.

Who this is actually for

  • Singers who want to practice intonation on a favorite song without buying stem software.
  • Home recordists cleaning up hiss or hum on a dry vocal before sharing a demo.
  • Students comparing pitch traces on an isolated estimate vs the full mix.
  • Anyone curious what an acapella-style pass sounds like, with a free download and no account.

Limits

  • Clips shorter than 3 seconds are too short to separate.
  • On phones the model runs on one worker and is slower; if the browser runs out of memory, noise reduction may still work.
  • Lead and backing vocals are not split from each other.

Credits

Separation model: KUIELab-MDX-Net (Leaderboard B vocals) by Minseok Kim, Woosung Choi, Jaehwa Chung, Daewon Lee, and Soonyoung Jung (2021), Zenodo, doi:10.5281/zenodo.6513062. Weights under CC BY 4.0, code under the MIT license. Runs with ONNX Runtime Web.