WAV vs MP3 for vocal separation: does the format actually matter?

May 2026 · 6 min read

People ask us this constantly: "I only have a 320 kbps MP3 — will separation still work?" Usually, yes. Sometimes, noticeably worse. The gap is smaller than audiophile forums suggest and larger than free-tool landing pages claim.

What the model actually sees

AI stem separators like the Spleeter model we run on Replicate do not magically reconstruct lost frequency data. They analyze patterns in the spectrogram — vocal formants, drum transients, stereo width — and guess which energy belongs to which stem. A lossy codec has already thrown away information the model might have used to draw cleaner boundaries.

That loss is cumulative. A WAV ripped from CD, processed once, uploaded once? Fine. An MP3 that was downloaded from a repost account, trimmed in a phone app, and re-encoded twice? The model is working with leftovers.

Where MP3 hurts most

Low bitrates (128 kbps and below) smear consonants and collapse stereo detail. Vocals that sit wide in the mix get harder to isolate because left/right differences are partly artifact, partly signal. Heavy limiting on a mastered MP3 can also confuse the separator — everything is loud, so everything looks "important."

Genres with dense high-frequency content — cymbal-heavy rock, bright pop, some EDM — show MP3 artifacts faster than sparse acoustic tracks. If you hear swishy hats before separation, MP3 encoding may be part of the problem.

When MP3 is good enough

A clean 256–320 kbps MP3 from a legitimate source is often indistinguishable from WAV for separation purposes, especially on pop and hip-hop with centered vocals. We accept MP3, WAV, M4A, and FLAC uploads up to 15 MB. For a three-minute track, that size cap rarely forces you into a bad MP3 — but a four-minute FLAC might need a downsample or shorter clip.

Practical rule: if the original sounds good to you in headphones, the format is probably fine. If it already sounds thin or crunchy, separation will not fix that.

Export format after separation

SongRemoveVocals lets you download instrumental and vocal stems as WAV or MP3. WAV is the right choice if you are dropping the track into a video editor or DAW for further EQ. MP3 is fine for a quick karaoke rehearsal or a draft cut. Re-encoding an already lossy source to MP3 again only adds another generation of damage — export WAV when you might edit later.

A fair test on your material

The honest comparison is A/B on the same song. Upload a WAV version and an MP3 version of identical content to the tool, listen to the chorus on both instrumentals, and pick the winner. Ten free minutes per day (UTC reset) is enough for a couple of test passes. For a broader workflow overview, see our how-to guide.