VocalSplitv1.0.0

Separates songs with the Demucs model on your own computer into an instrumental and an a cappella, or into drums, bass, other, guitar and piano stems, from MP3, WAV, FLAC, M4A or video.

See what you can make

Made with this SkillInstrumental, vocals, stems and custom mixes with a mixer to compare them

Vocals 17.1 dB SDR against the true part

Open full example

What you get

Instrumental, vocals, stems and custom mixes with a mixer to compare them

What you can make

  • Separates songs with the Demucs model on your own computer into an instrumental and an a cappella, or into drums, bass, other, guitar and piano stems, from MP3, WAV, FLAC, M4A or video.
  • Makes practice mixes with a quiet guide vocal and karaoke versions in another key, keeps every file adding back up to the original, and lowers all files together if one would clip.
  • Measures loudness and peaks, finds where the vocals are, and builds an offline mixer page with every part in sync, mute, solo, levels and A/B.
Explore delivered files6 groups
  • instrumental: the karaoke track
  • vocals: the a cappella
  • stems/: drums, bass, other (guitar and piano with htdemucs_6s)
  • mixes/: practice mixes and key changes
  • index.html: offline mixer with mute, solo, levels and A/B
  • report.md with loudness, peaks and where the vocals are

Good fit

Singers, karaoke hosts, musicians practising, DJs and producers, and video makers working with their own or licensed music.

Know the limits

Separation is an estimate: expect faint vocal traces and some band in the vocal, especially from MP3s. Needs PyTorch and Demucs installed; it does not clear rights to commercial recordings.

About this skill

Your agent checks that the recording is yours or licensed, installs Demucs and PyTorch once, and writes the settings. VocalSplit decodes MP3, WAV, FLAC, M4A, OGG or the audio of a video with FFmpeg and separates it with Demucs: htdemucs for speed, htdemucs_ft for the cleanest vocals or htdemucs_6s for guitar and piano too, on an NVIDIA or Apple GPU when present or the CPU. The instrumental is the song minus the vocals, so vocals and instrumental add back up to the original exactly, and a null test confirms it. Optional stems, practice mixes with any level per part (a guide vocal at 25%, say) and karaoke versions shifted up to 12 semitones with the tempo kept are written as WAV, 24-bit WAV, FLAC or MP3 at the source's sample rate; if any file would clip, all are lowered together. Each file's loudness and true peak are measured, the vocal sections are found, and an offline mixer page plays every part in sync with mute, solo, levels, A/B and a vocal timeline. Requires Python 3.9+, PyTorch, Demucs and FFmpeg.

FROM PACKAGE TO FIRST RESULT

Your first result, with your agent

Choose a starting point, add the content you already have, then let your agent check the setup and create one result.

View example result

Purchase access is required. Unzip the package into the project folder below, or let your agent follow the installation guide. .agents/skills/vocalsplit

Your project name, audience and the content or images you already have are enough to start. Your agent will ask for anything essential that is missing.

3. Let your agent guide the first result

Package location: .agents/skills/vocalsplit

Use VocalSplit to make a karaoke instrumental of my band's song. Start from examples/vocalsplit.json.

Use a new output folder and preserve existing files. Inspect the actual deliverables (instrumental: the karaoke track, vocals: the a cappella, stems/: drums, bass, other (guitar and piano with htdemucs_6s)), open the visual result and explain failed checks. An incomplete dependency check is not a successful result. Ask before paid services and do not publish automatically.
Before calling it done
  • Dependencies checked before rendering
  • Actual files open and work with your content
  • Editable source saved in a new folder
Made your first result? Share your experience →
Installation options & technical requirements

Start with your agent

Give this request to Claude Code, Codex or Cursor. Your agent can collect the missing details and guide your first result.

Use VocalSplit to make a karaoke instrumental of my band's song. Start from examples/vocalsplit.json.

Purchase access is required to install the package. Technical commands are available below when you need them.

US$4.99 · MIT licensePython 3.9+ · PyTorch and Demucs · FFmpeg · runs locally, no API keyNo API key or paid provider
  1. 01

    Choose your agent

    These steps use a project folder. Open the project in your agent and run terminal commands from its root.

    Codex directory guide

    Package folder .agents/skills/vocalsplit

  2. 02

    Add the package

    Download and unzip the package into the folder above. Keep all its files together, with SKILL.md directly inside that folder.

    Downloads require a verified account and purchase access. Choose a new folder if this Skill is already installed.

  3. 03

    Run the included sample

    Manual sample commands and technical details

    Read SKILL.md, review the generator and check its runtime requirements before running:

    cd .agents/skills/vocalsplit
    pip install -r requirements.txt
    python scripts/vocalsplit.py run --input examples/vocalsplit.json --output ./separated
    Your first result

    Open index.html, listen to the instrumental through the loudest chorus, and use the files you need.

    The demo song is original, made with a synthesized band and FFmpeg's flite voice. Separation is an estimate; use it only on recordings you have the rights to.

    If separated already exists, change the output to ./separated-2. Existing files are never overwritten.

    Prefer to ask your agent?

    Use VocalSplit to remove the vocals from my song: separate it with Demucs into an instrumental and an a cappella vocal (and stems if I ask), make a practice mix or a key change if I want one, and give me the mixer page to listen before I use them.

Reviews

Share your first result

Tried this workflow? Help others decide with an honest review.

Before you choose this Skill

How do I get VocalSplit?

The listed individual package price is USD 4.99. Premium downloads require a confirmed purchase or active membership. Your agent subscription, model access and external provider credits are separate.

Which agents can use this skill?

This release lists Codex, Claude Code, Cursor as intended installation hosts. Check the package’s dependencies and permissions before running it. These labels do not certify that every workflow has been tested in every agent.

What files can this workflow generate?

The documented 1.0.0 workflow lists: instrumental: the karaoke track; vocals: the a cappella; stems/: drums, bass, other (guitar and piano with htdemucs_6s); mixes/: practice mixes and key changes; index.html: offline mixer with mute, solo, levels and A/B; report.md with loudness, peaks and where the vocals are. Some files or exports are optional; follow the package’s setup and inspect the delivered result.

What should I check before using the result?

The demo song is original, made with a synthesized band and FFmpeg's flite voice. Separation is an estimate; use it only on recordings you have the rights to.

$4.99One-time purchase