GitHub - Siziff/musslop: Any track becomes a game soundtrack: AI-detected sections, seamless loops, beat-aligned transitions, intensity layers.

Any track becomes a game soundtrack

Loop any section forever · hit «Next» · the music evolves — seamlessly, on the beat

In games, music reacts: stay in the tavern and its theme loops forever; descend into the dungeon and the score darkens with you. musslop does this with any mp3.

Drop a track → AI finds the musical structure → every section becomes a perfect loop → you drive the music live, like a game audio engine — no editing skills required.

Why it feels magic

Feature	What it does
AI structure analysis	two neural engines find sections and name them (Intro, Verse, Chorus, Solo): SongFormer (2025, faster, best on pop/rock/electronic) and All-In-One (2023, steadier on orchestral); beats/downbeats via Beat This!. Fast heuristic fallback (~2 s) works everywhere
Seamless loops	sample-accurate Web Audio scheduling, bar-snapped boundaries, per-section loop-quality score (⟳%), equal-power crossfades
Intensity layers	Demucs splits the track into drums / bass / vocals / backing — toggle and mix layers live: calm exploration → full combat, same track
Pro transitions	post-exit tails ring out over the next section, bass-swap keeps exactly one bassline at any moment, one-shot stingers (cymbal / boom / riser) punctuate scene changes
Knows what not to loop	build-ups are detected by their crescendo shape and play once, as dramatic bridges
Full editor	drag boundaries with bar snapping, split/merge, loop-start markers, Ctrl+Z, zoom + scrollbar, everything auto-saved

Built for the game table

The flagship use case is tabletop RPG: turn any track into a location theme. The party lingers in the tavern — the tavern section loops. They open the dungeon door — one keypress, and the music descends with them, on the beat, mid-session. There is even a Tavern UI theme (wood, parchment & gold) switchable in the header.

Also great for: game dev prototyping (audition adaptive behaviour before wiring FMOD/Wwise, export loop WAVs), streaming, practice looping, focus music.

Quick start

Linux / macOS:

git clone https://github.com/Siziff/musslop.git && cd musslop
./setup.sh   # creates .venv, installs everything (incl. a bundled ffmpeg)
./run.sh     # → http://localhost:8801

Windows (needs Python 3.10+ with "Add to PATH" checked):

git clone https://github.com/Siziff/musslop.git
cd musslop
setup.bat
run.bat      # → http://localhost:8801

ffmpeg is downloaded automatically with the dependencies — no manual install needed.

Optional AI engines (each is one command, both auto-detected by run.sh):

Linux / macOS:

./setup-ai-songformer.sh  # SongFormer 2025 + Beat This! — newer, faster,
                          # best on pop/rock/electronic; plain pip, no compilers
./setup-ai-allin1.sh      # All-In-One 2023 + Demucs — steadier on orchestral,
                          # also powers the stem "layers" feature

Windows:

setup-ai-songformer.bat   # needs git in PATH; ~15 min, ~4 GB

The All-In-One engine is not supported natively on Windows (its NATTEN dependency needs a source build); use WSL2 with the Linux script if you need it. SongFormer covers structure analysis fully, and the stem "layers" feature also works through it on Windows.

Works fully offline after setup. First AI analysis of a track: ~1 min on GPU, a few minutes on CPU; results are cached — reopening is instant.

Under the hood

Beat & downbeat tracking — onset envelope + dynamic programming (Ellis 2007)
Structure — neural (All-In-One, WASPAA 2023) or Foote novelty on beat-synced chroma+MFCC+RMS self-similarity
Loop-aware refinement — boundaries move along downbeats maximizing loop closure quality (head/tail spectral similarity), phrase lengths (4/8 bars) and transition audibility
Build-up detection — normalized RMS slope + trend R² + spectral-centroid rise
Playback — every loop pass is an independent AudioBufferSourceNode; transitions land on loop/phrase boundaries with micro-fades, tails and bass-swap

Built on the shoulders of: SongFormer (ASLP-lab, CC-BY-4.0) · All-In-One (Kim & Nam, MIT) · Beat This! (CPJKU, MIT) · Demucs · librosa — plus classic MIR: Ellis 2007 (beat tracking), Foote 2000 (novelty segmentation).

API

Endpoint	Purpose
`POST /api/upload` · `POST /api/import_url`	file upload / yt-dlp import
`GET /api/analyze/{id}?engine=fast\|deep`	structure analysis
`POST /api/stems/{id}` · `GET /api/stems/{id}/{stem}`	Demucs 4-stem split
`POST /api/export/{id}`	zip of full-quality loop WAVs
`GET/POST /api/markup/{id}`	segment markup persistence
`GET /api/tracks` · `POST /api/favorite/{id}`	history & library

Troubleshooting

curl http://localhost:8801/api/health → expect {"status":"ok", ...}
Blank page → hard-refresh (Ctrl+F5), check browser console (F12)
Port busy → ./run.sh frees it; manually: lsof -ti tcp:8801 | xargs kill
No sound → click the ♪? self-test in the volume box (bottom-right)
Server errors → server.log