What is noCap?
noCap takes an audio file, isolates the vocals, transcribes every word with
word-level timestamps, and aligns each syllable to a beat grid. The result is
an interactive flow map — a piano-roll-style canvas where you can see and
hear exactly how a rapper rides the beat.
It answers questions like:
- How dense is the flow? (syllables per beat)
- How much syncopation is there?
- Is the delivery ahead, on, or behind the pocket?
- Where do the rhyme chains land on the grid?
You can also read the whole track as a lyric sheet, fix any word it misheard
(the flow map and metrics update with it), and export the lyrics as text or
synced .lrc.
Flow is the part of rap that is hardest to put into words. Producers, rappers and
fans describe it by ear: "laid back", "on top of the beat", "double time". noCap
makes it something you can point at, bar by bar.
Install
Everything noCap needs is bundled, and analysis runs offline on your computer.
The first time you open it, a short intro shows how to read the flow map; drop in
an audio file, or press Try the demo to explore a sample verse first.
macOS
- Download the latest
.dmg from Releases.
- Open it and drag noCap into Applications.
Needs a Mac with Apple Silicon (M1 or later). The app is signed and notarized by Apple.
Windows
- Download both
noCap-<version>-win-x64.7z.001 and .7z.002 from
Releases into the same folder
(the app is split in two because of GitHub's file size limit).
- Open the
.001 file with 7-Zip and extract it.
- Run
noCap.exe from the extracted folder.
Needs 64-bit Windows 10 or 11. An NVIDIA graphics card speeds up analysis; without
one it runs on the processor, just slower. The Windows build isn't code-signed yet,
so SmartScreen may warn you the first time: choose More info → Run anyway.
Features
- Beat detection — BPM and beat grid from the raw waveform (librosa)
- Vocal separation — Demucs stem isolation for cleaner transcription
- Speech recognition — Whisper
medium.en (MLX on Apple Silicon) with word-level timestamps
- Cleaner lyrics — filters for Whisper hallucinations and a second pass over passages it skipped; measured on a 20-song benchmark (WER 33.9 % → 27.6 %)
- Syllable alignment — CMU Pronouncing Dictionary for stress patterns
- Flow metrics — density, syncopation, pocket, timing variance per bar
- Rhyme detection — phonetic matching across perfect, slant, and internal rhymes
- Interactive visualization — WebGL canvas (Pixi.js), hover tooltips, click-to-play
- Lyrics view — the whole track as a lyric sheet with rhyme colours; fix misheard words and the analysis updates; export
.txt, synced .lrc or word-timed .lrc
- Live progress — plain-language steps, time left and small findings while a track is analysed; cancel at any point
- Library — save and re-open analyzed tracks; a demo verse and a short first-launch intro
- Desktop app for macOS and Windows — self-contained Electron + PyInstaller bundle (no setup required)
Keyboard shortcuts
Press ? in the app to see this list at any time. Shortcuts are ignored while you are typing in a field or while a dialog is open.
| Group | Keys | Action |
|---|
| Playback | Space | Play / pause |
| Playback | Enter | Play selected bars |
| Playback | L | Loop selected bars |
| Playback | V | Switch mix / vocals |
| Navigation | ← → | Previous / next bar |
| Navigation | ⇧ + click | Extend selection |
| Navigation | ↑ ↓ | Scroll |
| Navigation | Esc | Clear rhyme focus, then selection |
| View | ⌘ + | Zoom in |
| View | ⌘ − | Zoom out |
| View | ⌘ 0 | Reset zoom |
| View | I | Show / hide inspector |
| View | 1 2 3 4 | Flow / Report / Lyrics / Library |
| View | ? | This list |
On Windows and Linux, use Ctrl in place of ⌘.
Getting Started
Prerequisites
Development
# 1. Clone
git clone https://github.com/fatihserhatturan/noCap.git
cd noCap
# 2. Python environment
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,separator]"
pip install openai-whisper
pip install -e ".[apple]" # Apple Silicon only: GPU transcription via MLX
# 3. Frontend dependencies
cd frontend && npm install && cd ..
# 4. Run (Flask API + Vite dev server + Electron window)
cd frontend && npm run desktop:dev
The first time you analyze a track, Whisper will download the medium.en
model (~1.5 GB) into ~/.cache/whisper/. Subsequent runs use the cache.
On Apple Silicon, analysis runs on the GPU: Demucs uses MPS and, with the
apple extra installed, Whisper runs through MLX from the same checkpoint (same
DTW word timestamps, beam search). Anything that fails on the GPU falls back
to the CPU. NOCAP_WHISPER_BACKEND=mlx|torch|cpp and NOCAP_DEMUCS_DEVICE=cpu
pin a backend or device.
Web-only mode (no Electron)
# Terminal 1 — backend
source .venv/bin/activate && nocap serve
# Terminal 2 — frontend
cd frontend && npm run dev
Open http://localhost:8765.
Building the macOS App
Produces a self-contained .dmg that requires no Python or model installation
on the target machine.
./scripts/build-desktop.sh
# → frontend/dist-app/noCap-0.3.0-arm64.dmg (Apple Silicon)
First build takes ~15 minutes (PyInstaller bundles torch + whisper + demucs).
Subsequent builds that skip the backend are much faster:
./scripts/build-desktop.sh --skip-backend
Signing. When the Keychain holds a Developer ID Application certificate and a
notarytool profile named nocap-notary
(xcrun notarytool store-credentials "nocap-notary" --apple-id … --team-id …),
the build signs, notarizes and staples the app, so Gatekeeper opens it without a
warning. Without them it falls back to an ad-hoc signature (users confirm the app
once in Privacy & Security); --ad-hoc forces that fallback.
Where Your Data Lives
The library is one SQLite database plus a folder of audio files, in the standard
per-user app-data location:
| Run | Folder |
|---|
| Installed app | ~/Library/Application Support/noCap/ |
Development (nocap serve, npm run desktop:dev) | ~/Library/Application Support/noCap-dev/ |
library.db tracks, analyses (compressed flow maps), bars, sections
audio/<track_id>/ mix.<ext>, vocals.wav
Set NOCAP_DATA_DIR to use another folder. To back up the library, copy the whole
folder while the app is closed; to reset it, delete the folder.
Tracks saved by earlier versions (~/.nocap_library/tracks) are copied in
automatically on first start; the old folder is left untouched and can be deleted
afterwards.
Tech Stack
| Layer | Technology |
|---|
| Desktop shell | Electron 33 |
| Frontend | React 19, TypeScript, Pixi.js 8, Vite 6 |
| Backend | Python / Flask 3 |
| Library storage | SQLite (stdlib sqlite3, WAL) |
| Beat detection | librosa 0.11 |
| Transcription | openai-whisper (medium.en), MLX on Apple Silicon |
| Vocal separation | Demucs 4 (htdemucs) |
| Syllable/stress | CMU Pronouncing Dictionary |
| Packaging | PyInstaller + electron-builder |
Project Structure
nocap/
├── frontend/src/ # React UI — components, hooks, Pixi rendering
├── nocap/audio/ # Beat detection, transcription, vocal separation
├── nocap/analysis/ # Flow metrics, alignment, rhyme detection
├── nocap/pipeline/ # Orchestration layer
├── nocap/web/ # Flask API + SSE streaming
├── nocap/text/ # Lyric parsing, syllable analysis
├── nocap/library/ # Track persistence
└── scripts/ # Build automation
All source files are kept under 200 lines — see CODE_QUALITY.md.
Development Checks
python3 -m unittest -q
python3 scripts/check_line_limits.py
cd frontend && npm run typecheck && npm run build
Contributing
Bug reports, ideas and pull requests are welcome.
- Found a problem or have an idea? Open an issue, or start a thread in Discussions.
- Want to write code? Issues labelled
good first issue are a good place to start. See Development to run it locally, and CODE_QUALITY.md for the house rules.
Roadmap
- Builds for Intel Macs and Linux
- A signed Windows installer
- Better support for lyrics in languages other than English
- Comparing two verses or two rappers side by side
- More export formats for the flow map and the report
If noCap is useful to you, a star helps other people find it.
