Practice Stems

Moises-style practice app for Mac/Linux: YouTube or local audio/video → Demucs 6-stem split → live mixer (volume / mute / solo).

Stems: vocals, drums, bass, guitar, piano, other (htdemucs_6s).

Screenshots

Home — paste a YouTube URL or upload a file, then open songs from your library:

Stem mixer — Play / seek, role presets, and per-stem volume / mute / solo:

Presets cycling (Guitarist, Drummer, Singer, …):

---

Option A — Local with conda (recommended on Mac: Apple GPU)

Faster than Docker on Apple Silicon. Needs Miniconda/Anaconda and Homebrew ffmpeg/rubberband only if conda packages are missing.

cd practice_stems
conda env create -f environment.yml
conda activate practice-stems
python app.py

If the env already exists and you only need to refresh deps:

conda activate practice-stems
pip install -r requirements.txt

Open http://127.0.0.1:7860

Needs: network for YouTube + first model download. For HQ slowdown: brew install rubberband (CLI on PATH).

---

Option B — Docker (required / portable)

1. Start a Docker engine

Colima (common on Mac without Docker Desktop):

brew install colima docker
colima start --cpu 4 --memory 8

Use at least 8 GiB RAM. With 2 GiB the container will die during separation (OOM).

Check:

colima list

Or start Docker Desktop and wait until it’s running.

2. Run the app

cd practice_stems
./run-docker.sh

• If an image already exists, choose Y to reuse it (app code is bind-mounted from this folder).
• Choose n to rebuild the image from scratch.
• If Colima RAM is too low, the script can restart Colima with 4 CPUs / 8 GiB.

Open http://127.0.0.1:7860 — stop with Ctrl+C.

Equivalent manual commands:

docker build -t practice-stems .
docker run --rm -p 7860:7860 \
  -w /app -e PYTHONPATH=/app -e HOST=0.0.0.0 \
  -v "$(pwd)/data:/app/data" \
  -v "$(pwd)/app.py:/app/app.py:ro" \
  -v "$(pwd)/pipeline:/app/pipeline:ro" \
  -v "$(pwd)/static:/app/static:ro" \
  -v "$(pwd)/.cache/torch:/root/.cache/torch" \
  practice-stems

Docker notes

Topic	Detail
Songs data	./data on your machine → /app/data in the container
Model cache	./.cache/torch (avoids re-downloading Demucs weights)
GPU	Docker on Mac = CPU only. Linux + NVIDIA can use CUDA later
First separate	Slow: downloads model (~50 MB) then processes the whole track
Full songs	Leave the terminal open until progress finishes

---

How to use the app

1. Paste a YouTube URL, or upload audio / video, then click Open song (download/extract only — no Demucs yet).
2. On the song hub: preview the track, then pick an analysis:
   • Stem separation → opens the practice mixer when ready
   • Detect BPM → tempo estimate
   • Detect chords → madmom DeepChroma (maj/min); prefers guitar+piano+other stems
3. In the mixer: Play, Speed, Key −/+ (shows key name e.g. Bm→Am; change anytime while playing), HQ speed On/Off, faders, Mute / Solo, presets, Download mix / Download stems.
   • Prefer 0.75×–0.9× for practice; turn HQ speed On when quality matters.
   • On the song hub, Key −/+ shifts chord labels + preview audio (key name from Detect chords / auto detect).
4. Library Open returns to the song hub (not straight to stems).
5. Stop loading cancels an in-progress download/open.
6. Delete removes a song from the library.

---

Data layout

practice_stems/data/
  play/          # library + MP3s for the mixer
  _incoming/     # downloads + Demucs WAV stems
  _uploads/      # temporary uploads

---

Troubleshooting

Problem	Fix
docker.sock: no such file	Start Colima (colima start) or Docker Desktop
docker compose unknown	Use ./run-docker.sh (this install has no Compose plugin)
Container dies mid-separate	Raise Colima RAM: colima stop && colima start --cpu 4 --memory 8
TorchCodec errors	Fixed in current code (uses soundfile + ffmpeg). Restart container / remount
Shape / reshape errors	Fixed (no custom Demucs segment). Restart ./run-docker.sh so mounts pick up latest pipeline/
Port 7860 in use	Stop the other app (Ctrl+C) or change host port: -p 7861:7860

---

Requirements files

• requirements.txt — local Mac/Linux (includes current torch)
• requirements-docker.txt — Docker image deps (CPU torch installed separately in the Dockerfile)