Skip to content
TimeCues Studio

TimeCues

TimeCues Studio is a web app for annotating musical structure and evaluating boundary-detection algorithms. Annotators mark section boundaries on a waveform; the app overlays algorithmic estimates from allin1, MSAF, ruptures, and change-point detectors; mir_eval runs the metrics in-browser.

Open source, multi-arch (amd64 / arm64), with a CI pipeline and a public live demo.

What it does

  • Annotate. Lay down section boundaries on a per-song waveform with keyboard-driven flow. Multiple layers per song — manual (reviewed, source of truth), auto-guess (algorithm-clustered draft you tick through point-by-point).
  • Compare. Toggle algorithm overlays. Each algorithm’s boundary set gets its own color and a per-song F-measure / HitRate against the current manual layer.
  • Evaluate. Run mir_eval metrics across a whole dataset in one click. Aggregate scores per algorithm and per annotator.
  • Multi-annotator. Each annotator’s work lives under their own subdir. Sign in with Google; compare annotations side-by-side; resolve disagreements.
  • Custom detectors. Drop a Python file in tools/python/custom/, reload, and your algorithm shows up as another overlay with full evaluation. No build step.

Where to go next

  • User guide — every panel, button, shortcut, and file format. Start here if you want to use the app.
  • Data model — how songs and annotations are laid out on disk. Read if you’re integrating with the data directly.
  • Experimental models — opt-in detectors behind feature flags. May break.
  • Deployment — put it on a public URL: a cloud VM running Docker Compose, with HTTPS and Google sign-in. Read if you want your own shared instance.

Try it

If you’d rather poke at the running instance than read docs, the public deployment lives at the address in the project README. Sign in with Google; you’ll start in your own namespace.

The paper

There’s an accompanying paper describing the annotation tool, the evaluation methodology, and findings on inter-annotator agreement. The PDF is linked from the project’s GitHub README.