What 128 kbps and 320 kbps change in the split
An MP3 is already a reduced copy of the mix. The separator never sees the session tracks. It sees whatever survived the encode. Bitrate is the budget for that survival, and it shows up in the stems even when both files are the same song.
At 128 kbps, a typical encoder spends its bits on the loud middle of the spectrum and applies a lowpass well below 20 kHz. The exact cutoff depends on the encoder, not on this site. Consonants, breath, and the top of a hi-hat are the first things to thin out. Those are also the cues a vocal model uses to decide “this transient belongs to the singer” versus “this transient belongs to the kit.” With fewer of those cues, the instrumental tends to come back duller, and the vocal stem tends to keep a soft, unstable edge.
At 320 kbps the same master keeps more of the top octave and less quantization noise. The split is still an estimate from a stereo mix, not a studio stem, but the model has a cleaner picture of sibilance and cymbal energy. If you have a choice of MP3s, upload the 320 kbps file. If 128 kbps is all you have, the tool will still run. Expect the difference in the vocal edge, not in whether the singer disappears.
How MP3 damage shows up as a watery vocal
MP3 coding does not only delete quiet highs. It also spreads sharp attacks slightly in time, which engineers call pre-echo, and it leaves a low-level noise floor shaped like the music. A separator trained to pull a singer out of a mix can treat that noise as part of the voice, because it sits in the same band as breath and room tone.
On the vocal stem that often sounds watery: a swirl around the note, a soft doubled edge, or a hint of the hi-hat that will not sit still. It is not a separate reverb plugin, and it is not something a second pass of the same MP3 will wash off. The vocal remover writes two MP3s from the upload it received. If the source already had that smear, both stems inherit a share of it.
A second generation makes this worse. An MP3 ripped from a video, then exported again from a phone editor, has been through the lowpass twice. Going back to the first encode, or to a lossless file if you still have one, changes the input. Re-exporting the same MP3 at a higher bitrate does not.
When it is worth converting to WAV first
Convert to WAV when the WAV is a better recording than the MP3. A session bounce, a CD rip, or a FLAC you still have contains highs and transients the MP3 never stored. Upload that file. The free tier accepts WAV, and a two-minute 16-bit, 44.1 kHz stereo WAV is about 20 MB, under the 50 MB cap.
Most of the time the conversion is not worth it. If the only copy you have is an MP3, decoding it to WAV and uploading the WAV gives the model the same samples it would have decoded from the MP3 itself. The missing band does not come back. This site accepts MP3 directly, so the extra transcode is just a larger file and a chance to clip the bounce if the export is careless.
The useful order is the opposite of a ritual convert-everything step: keep the earliest, least-encoded file you are allowed to use, trim it to the section you need, and upload that. The free workflow on the no-signup page is the same either way.
How long an MP3 can be before 50 MB
The upload cap is 50 × 1024 × 1024 bytes, about 52.4 million bytes. At a constant bitrate, duration in seconds is that size times 8, divided by the bitrate in bits per second.
| File | Approximate length at the 50 MB cap |
|---|---|
| 128 kbps MP3 | About 54 minutes 37 seconds |
| 320 kbps MP3 | About 21 minutes 51 seconds |
| 16-bit 44.1 kHz stereo WAV | About 4 minutes 57 seconds |
The free tier never lets you use those full lengths. Anything over 2 minutes needs Credits before separation starts. Two minutes of 128 kbps MP3 is about 1.8 MB. Two minutes of 320 kbps MP3 is about 4.6 MB. For a normal stereo MP3, the clock runs out long before the byte cap.
Size starts to matter when the file is not a plain MP3: a 24-bit or 96 kHz WAV, or a multichannel bounce, can cross 50 MB inside a short duration and fail at upload. Trim to the chorus or the verse you actually want to hear, then upload that section to the karaoke maker.