The most direct way to make a karaoke track, practice a vocal, or keep the backing music for a video is to separate a song into vocals and instruments, then leave out the vocal stem. You can now do that in a browser without installing a large audio application.
ToolGarden uses a browser-local HT-Demucs model to estimate the sources in a song. The selected audio is not sent to a remote separation API. The first use downloads the model, while decoding, inference, and WAV export run on the current device.
Are removing vocals and extracting an instrumental the same thing?
Vocals, drums, bass, and instruments in a finished song have already been mixed into the same waveform. Removing vocals does not delete an existing channel. A model estimates a vocal stem, and the remaining sources are used as the instrumental.
A two-stem tool normally outputs vocal and instrumental directly. A four-stem tool outputs vocals, drums, bass, and other, giving you separate control over rhythm and low end. ToolGarden also has a six-stem model that adds guitar and piano.
Should you use four stems or six stems?
| Mode | Output | Best use | Recommendation |
|---|---|---|---|
| Four stems | Vocals, drums, bass, other | Vocal removal, karaoke, singing practice | Use this first for a cohesive instrumental |
| Six stems | Vocals, drums, bass, guitar, piano, other | Independent guitar or piano control | Choose only when those instruments are needed separately |
For a complete backing track, start with four stems and keep drums, bass, and other. Six stems do not automatically produce cleaner audio. They divide the same mix into more categories by pulling guitar and piano out of other.
Steps to remove vocals online
- Prepare the best available source, ideally WAV, FLAC, or a high-bitrate MP3.
- Open the stem splitter and click the upload area or drag the song onto the page.
- Choose the four-stem model and select drums, bass, and other. Select vocals too if you want to inspect the separation.
- Start separation, keep the page open, and wait for model loading and audio processing.
- Preview drums, bass, and other to check for obvious gaps or artifacts.
- Download the three WAV files and align them at the same start time in an audio editor, or download all stems for later adjustment.
How to recombine the result into a full instrumental
In four-stem mode, the instrumental consists of drums, bass, and other. Those WAV files share the same start point, duration, and sample rate. Align them at the same timeline position in an editor or DAW to reconstruct the song without the lead vocal.
- Drums carry the beat and most transients, so do not reduce them heavily by default.
- Bass provides low-end weight; check that it is active when the instrumental sounds thin.
- Other normally contains guitars, piano, synths, strings, and sometimes backing vocals.
- Do not time-stretch the stems independently or they may drift out of alignment.
- Leave headroom before export so summing several stems does not clip.
How to get a cleaner instrumental
- Use a source that has not been transcoded repeatedly. Low-bitrate MP3 artifacts can be amplified by separation.
- Test a 30–60 second chorus first because it usually contains vocals, drums, bass, and a dense arrangement.
- Prefer four stems for a backing track to avoid unnecessary splitting between similar instruments.
- A little vocal residue can sound more natural than aggressive cleanup that damages snare, guitar, and high frequencies.
- Listen on headphones and speakers to check both centered vocal residue and low-end continuity.
Why can you still hear some vocals?
The model works from a finished mix rather than original studio sessions. Vocal reverb spreads across the stereo field, harmonies may share frequencies with synths, and distorted guitars can resemble vocal harmonics, so small amounts of bleed are difficult to eliminate.
Residue limited to reverb tails normally does not disrupt singing practice. If the entire lead remains obvious, try a higher-quality source, compare four- and six-stem output, or retest the specific section.
Common problems and fixes
| Problem | Likely cause | Fix |
|---|---|---|
| The instrumental sounds hollow | Bass or other was omitted | Combine drums, bass, and other |
| Lead vocal remains obvious | Reverb, harmonies, or overlapping frequencies | Use a better source and compare models |
| The rhythm echoes | Stem start points are misaligned | Start every WAV at the same timeline position |
| The export clips | Summed stems exceed peak level | Lower stem gain before export |
| The model will not load | Not enough available browser memory | Close other tabs and use desktop Chrome or Edge |
Can you publish or monetize the instrumental?
Removing a vocal does not change ownership of the recording or composition. Private practice, public performance, video uploads, cover releases, and commercial use can require different permissions. Process and publish only audio you own, are authorized to use, or may lawfully use.
Key takeaways
A good online instrumental workflow starts with the right separation mode. Use four stems for normal vocal removal, keep drums, bass, and other, and align them from the same start point. A clean source, correct alignment, and realistic expectations matter more than simply requesting more stems.