Same input for every tool · Unedited outputs · Measured and dated
Every vocal remover, same clip, side by side
One 50-second clip through AudioForges, Ultimate Vocal Remover, VocalRemover.org and LALAL.AI. Every output is playable below, unedited. Models, times and limits are printed with the date they were checked.
Vocals, every tool
Click a lane to switch while it plays. Drag on a lane to loop a section.
Press play. The waveforms load with the audio.
LALAL.AI has no download on its free plan, so its lane is a recording of the preview player, made 21 Sept 2026. Every other lane is the file the tool returned.
Method
How this comparison was run
The clip is a 50-second section of a real release with a lead vocal over a full arrangement. The exact same MP3 was uploaded to each tool, nothing was re-recorded or trimmed differently per tool, and each output above is what the tool returned, re-encoded to 192 kbps MP3 for the page. Where a tool names its model, the model is printed. Where it does not, that is printed too.
LALAL.AI is the one exception, and it is labelled. Its free plan plays a preview but does not let you download the result, so its lane is a recording of that preview player, made on 21 September 2026 with the Andromeda model selected. A recording can differ slightly from a paid download, so weigh that lane with that in mind.
On video
A second test, filmed
A separate run, recorded start to finish: 30 seconds of one song through VocalRemover.org, Moises, StemSplit, LALAL.AI and AudioForges, judged on sound, speed, the free plan and the price after it.
Watch the test
Measured
Models, times and limits
| Tool | Model | Time for this clip | Ran on | Free plan |
|---|---|---|---|---|
| AudioForges Standard | Forge 1 | 25 sec | GPU (RTX A5000) | Full song free to hear, no sign-up |
| AudioForges Clean | Clean, on Forge Studio 2 (band-split transformer, Sept 2026) | 53 sec | GPU (RTX A5000) | 30-second clip free, then 1 song |
| Ultimate Vocal Remover 5.6 | MDX23C-InstVoc HQ | 19 min 9 sec | i7-8750H laptop, CPU only | Free, desktop install |
| VocalRemover.org | Not disclosed | ~15 sec observed | Their servers | 1 free run per day |
| LALAL.AI | Andromeda (closed-source) | Not timed, preview recording | Their servers | Preview only, no download |
Times are end to end as a user experiences them, from submitting the clip to the result being playable, measured once each on 17 September 2026. LALAL.AI was not timed because its lane is a preview recording. UVR on a capable NVIDIA GPU is many times faster than the CPU figure shown; the laptop case is listed because that is the machine most people have. Try the same clip yourself on the AudioForges vocal remover.
Credit
The track
The clip is from What Would It Mean by H4RRIS feat. Nicole Apollonio, used with permission and gladly credited. Listen to the full track on Spotify or YouTube. Separating a song you did not make is fine for practice, sampling references and study; releasing the result needs the rights holder's permission, which is exactly what this page has.
Frequently asked questions
Judge with your own ears: every player on this page is the tool's unedited output from the same 50-second clip. The facts we can measure are printed in the table: which model each tool runs, how long the clip took, and what the free plan allows.
UVR runs on your own machine. On a laptop without a usable NVIDIA GPU it falls back to CPU, and MDX23C took 19 minutes 9 seconds on our i7-8750H for the 50-second clip. On a desktop with a good NVIDIA GPU, UVR is many times faster than that. The trade is the same either way: excellent free quality in exchange for the install, the model downloads and your own hardware.
A 50-second section of What Would It Mean by H4RRIS feat. Nicole Apollonio, used with the artist's permission and credited below. The exact same MP3 was uploaded to every tool: most on 17 September 2026, LALAL.AI on 21 September 2026.
LALAL.AI's free plan lets you listen to a preview but not download the result. To include it, the preview was recorded as it played, with the Andromeda model selected, and that recording is what you hear. Every other lane is the actual file the tool returned.