Stemmy

A local, GPU-accelerated stem-separation studio for Windows. Drop in a song or paste a YouTube link, separate it into vocals, bass, drums and sub-drums, guitar, piano, keys, and other stems, then mix, practice, identify songs, display lyrics, generate chords, tune an instrument, and export the results.

Audio separation, playback, mixing, analysis, tuner processing, and chord generation run on your machine. Shazam, lyrics, YouTube import, update checks, and model downloads use internet services only when requested.

What is new in v1.5

• Hidden app-style Windows launch with a dedicated maximized window, desktop shortcut, icon, and close-to-exit behavior.
• Background GPU protection so separation does not slow down when Stemmy is minimized or unfocused.
• Matched CUDA 12.8 stack for RTX 50-series systems: PyTorch 2.11.0, TorchVision 0.26.0, and TorchAudio 2.11.0.
• Complete install and repair workflow with source, dependency, CUDA, ONNX, startup, and shortcut verification.
• Local chromatic tuner and multi-genre Chord Creator.
• Basic Pitch ONNX MIDI/ASCII-tab export without TensorFlow.
• Improved Shazam/lyrics flow, persistent Karaoke sessions, retry-failed handling, safe package updates, Save As export, and live system monitoring.

See RELEASE_NOTES.md for the full v1.5 change list.

Features

• GPU stem separation: Quick, Standard, Deep, and optional Extended workflows.
• Deep drum breakdown: kick, snare, toms, hi-hat, ride, and crash.
• Full mixer: solo, mute, level, pan, waveforms, zoom, scrubbing, pitch, tempo, A-B looping, tap tempo, metronome, and per-stem export.
• Karaoke: playlist/single-link vocal removal, persistent sessions, retry failures, full-screen playback, synced lyrics, queue controls, album art, visualizers, and auto-advance.
• Song ID and lyrics: Shazam identification plus resilient LRCLIB searches, cleaned metadata, manual fallback, and saved-lyrics preservation.
• Musician tools: stable chromatic tuner, alternate tunings, local Chord Creator, MIDI transcription, and ASCII tab.
• Safe updates: individual update/rollback controls for yt-dlp, shazamio, and imageio-ffmpeg while the GPU/model stack remains protected.

Screenshots

Tested configuration

Component	Tested value
OS	Windows 11
GPU	NVIDIA GeForce RTX 5060 Laptop GPU, 8 GB VRAM
RAM	16 GB
Python	3.12.10
PyTorch	2.11.0+cu128
TorchVision	0.26.0+cu128
TorchAudio	2.11.0+cu128

Python 3.10–3.13 is supported; Python 3.12 is preferred. Budget roughly 15–25 GB free for the environment, model cache, and output stems.

Windows installation

Recommended: complete Windows release ZIP

1. Extract the complete Stemmy v1.5 Windows package into an empty writable folder.
2. Run install_all.bat.
3. Launch Stemmy from the desktop shortcut created by the installer.

GitHub source clone

Executable .bat launchers are intentionally excluded from Git through *.bat in .gitignore. This is deliberate: launchers belong in local folders and packaged Windows release assets, not in the source repository.

BUILD.md documents every launcher and links to the exact inert .bat.txt source under docs/launchers/. Save those files locally without the final .txt extension and use Windows CRLF line endings. Do not commit the recreated BAT files.

Upgrade or repair

1. Close Stemmy.
2. Extract the new package over the existing folder and replace source files.
3. Keep projects/, karaoke_jobs/, and models_cache/ when preserving work and downloaded models.
4. Run Repair Stemmy Installation.bat from the package or recreate it from BUILD.md.
5. Launch from the refreshed desktop shortcut.

The repair keeps the existing .venv whenever possible and avoids redownloading the large Torch wheel when only TorchVision needs repair.

Separation depths

Depth	Typical output
Quick	vocals, drums, bass, other
Standard	Quick plus guitar and piano
Deep	Standard plus detailed DrumSep and analysis
Extended	optional experimental MSST multi-instrument split

Quick, Standard, and Deep are part of the normal setup. Extended uses a separate large model and is enabled with locally recreated get_msst.bat.

Windows launcher and background GPU behavior

The desktop shortcut launches Stemmy without a visible PowerShell window. Python runs detached, logs are written under logs/, high-priority/high-QoS handling is applied to Stemmy and relevant child processes, and separation continues while the app is minimized or unfocused. Closing the dedicated app window requests a safe local shutdown.

Honest limitations

• Rhythm versus lead guitar, or clean versus distorted guitar, remains unreliable with current source-separation models.
• Detailed drum splitting is more dependable than same-instrument guitar splitting.
• Extended is experimental and RAM-heavy.
• Shazam, lyrics, YouTube import, update checks, and model downloads need internet access.
• YouTube changes can temporarily break downloads; update yt-dlp and retry failed Karaoke tracks.

Project layout

app/                       Flask app, pipeline, tools, Karaoke, lyrics, and UI
projects/                  separated output, ignored by Git
uploads/                   source audio, ignored by Git
karaoke_jobs/              saved Karaoke sessions, ignored by Git
models_cache/              downloaded models, ignored by Git
logs/                      runtime diagnostics, ignored except .gitkeep
docs/launchers/*.bat.txt   exact inert launcher documentation
*.bat                      local/release launchers, always ignored by Git

License

MIT. Third-party models and bundled libraries retain their own licenses.