How to choose a vocal remover model: compare the same passage
Compare one-minute vocal and instrumental samples. Learn what to listen for, when another model is worth trying, and how to choose a result that suits your project.
Start with the default Kim model. If you notice vocal bleed, damaged instruments or missing syllables, try another model on the same song and compare the troublesome passage.
1. Choose a song and listen to one minute
There are two songs here. The male-vocal example has Kim, SYH FT and SYH Big results; the female-vocal example currently has a Kim result only. Each excerpt is one continuous minute.
Listen to the original first, then audition the vocal or instrumental track on its own. Use the male-vocal example for comparisons between models; the female-vocal example shows how another recording separates.
Audition each track separately. Changing models keeps your position; changing songs starts from the beginning.
Click or drag a waveform to seek. On a focused waveform: ←/↓ back, →/↑ forward; Tab next track, Shift+Tab previous track. Space plays or pauses.
Loop
These examples are already processed. Listening uses no free allowance or credits.
Test settings and processing record
Tested on 7 November 2026: standard mode, 4-second segments, NVIDIA RTX 4060. Times cover local inference and export, excluding model loading, queue and transfers.
| Model | Processing time | Peak VRAM |
|---|---|---|
| Kim Mel-Band RoFormer | 7.63 s | 1,451 MiB |
| SYH Mel-Band RoFormer FT | 7.45 s | 1,449.2 MiB |
| SYH Big RoFormer V1 Fast | 7.64 s | 2,294.1 MiB |
The table follows the selected song. A single run does not establish a speed or sound-quality ranking. Previews are 192 kbps MP3.
Open the author’s vocal reference ↗
Reference recordings were aligned and level-adjusted. The male reference combines several author-supplied vocal tracks. They help identify the source part; they are not AI outputs or the original mix balance.
Full test record ↓2. What to listen for depends on the track you need
For karaoke or a backing track
Can you still make out lyrics in the instrumental? Have drum attacks, guitar or piano transients been dulled? Does the remaining lead vocal get in the way when you sing along?
For an a cappella or vocal edit
Are consonants, breaths and phrase endings intact? Does the voice sound thin, broken or uneven? Is there obvious music between vocal phrases?
One result may work better for singing along, while another preserves more vocal detail. When a passage sounds cleaner, also check that the sounds you want have survived.
If you hear a problem, compare that exact spot
- Recognizable singing in the instrumental: switch models at the same lyric, then check whether instruments have also been removed.
- Missing syllables or cut-off phrase endings: return to the original, then check how another model handles the same phrase.
- Softened drum hits or choppy instruments: compare their attacks and pauses with the original and decide whether the change affects your use.
3. Try the same method on your own song
- Upload audio you have the right to process. Once uploaded, choose a vocal/instrumental separation model.
- Run one model first. Check the output format and the quote before starting.
- Listen to the result and note a specific issue and timestamp, such as a clearly audible lead vocal at 1:12.
- Try another model if that issue matters. Keep the processing mode and listening position the same.
Processing again with another model creates a new task and uses the free allowance or credits you select. Replaying or downloading an existing result does not charge you again.
Need drums, bass or another instrument on its own? Choose the appropriate separation task first.
4. When can you stop comparing?
Judge the result in context: sing along to the backing track, or use the extracted vocal in the edit you have in mind. If one version meets that need and the others offer no meaningful improvement, save it.
This one-minute example teaches a comparison method. Your own song, a different vocal style or a denser arrangement may behave differently. There is no need to spend tasks chasing a single model for every song.
Audio sources and usage
Male vocal: It’s OVER!!! — Admiral Bob.
Female vocal: Music Through Love — Kara Square, featuring Admiral Bob. Includes source material from: Heavy Chill — Admiral Bob · That Song I Never Wrote — Admiral Bob.
The source recordings and listed samples use CC BY 3.0. Changes: continuous excerpts, vocal/instrumental separation and MP3 encoding.
If a sample affects your privacy or rights, contact support@uvr.online or open a Support ticket with the sample link to request a review and removal. Contact support →