English · 繁體中文
A single-file audio span annotator. Open earmark.html in a browser and mark time spans on a
waveform — tag them, label them, export them. Decoding, waveform, spectrogram, resampling and
export all happen in the tab: no audio leaves the machine, and there is no build step, no
server and no dependencies.
Built for preparing speech-enhancement eval data — marking which regions must survive processing and which must be attenuated — but it works as a general span annotator.
earmark.html — one file. Open it in a browser and you are done: no build step, no server, no dependencies, nothing to install.
- Load audio. Drag files or a folder onto the window, or use File ▸ Open folder….
A folder replaces the working set — it asks first if that would discard spans — while
dropping loose files adds to it. A
<name>.spans.jsonsitting next to an audio file is picked up automatically. - Mark a span. Drag a range across the waveform or spectrogram, then press Enter. It appears in the lane below and as a row in the table.
- Tag and label it. Pick a tag in the left rail before pressing Enter, or change it afterwards in the table. Type a free-text label in the Label column. Add your own tags at the bottom of the rail — they are remembered between sessions.
- Adjust. Drag a span's edges or its body in the lane, or edit the numbers in its row. ⌘Z undoes anything you do to spans.
- Move through the folder. [ and ] step between files. Each file keeps its own spans and its own undo history, so nothing is lost on the way.
- Export. Pick a format at the bottom of the span sheet and hit Download, or Preview
to open the result in its own tab. With a folder loaded, one export covers every annotated
file. On Chromium, Save all writes a
.spans.jsonsidecar next to each audio file.
The heading bar carries the readouts — file, selection, format, working rate, level — and the bottom bar carries only controls, so nothing reflows while you drag or play.
Beside the waveform the vertical axis is amplitude in dBFS; beside the spectrogram it is frequency in kHz, spanning the working rate's Nyquist.
Amp magnifies the waveform for quiet files. It is a display gain only — it never touches the
audio or the export — and the dBFS axis moves with it so the view stays honest. Boost is
playback gain behind a limiter, so it cannot clip; Auto sets the largest boost the file takes
with no limiting at all.
Copy image puts what you are looking at on the clipboard as one picture: waveform, spectrogram, the span lane with its labels, and both axes, captioned with the file and the visible time range. Where a browser refuses image writes it saves a PNG instead.
The swatch in the heading bar opens the theme and colour panel: auto / light / dark, plus five spectrogram ramps. Each ramp is monotonic in lightness and respaced at equal perceptual distance, so a step in energy always looks like the same size step.
| Space | play / pause | ⇧Space | play selection |
| Enter | add span from selection | 1…9 | pick tag |
| S / E | set start / end at playhead | ⌫ | delete selected span |
| ⌘Z | undo | ⇧⌘Z | redo |
| [ ] | previous / next file | L | loop selection |
| + − 0 | zoom in / out / fit | ← → | nudge 1 s (⇧ 0.1 s) |
| G | auto boost | \ | toggle the file list |
| ⌘S | save spans into folder | ⌘E | download export |
| ⇧⌘C | copy the view as an image | ? | the full list, in the app |
Files — single files, a whole folder, or drag-and-drop (folders included). Each file keeps its own spans. Pane sizes, the file-list/tag split and the rail width are draggable and remembered.
Sample rate — Playback ▸ SR defaults to Auto · file rate: each file is decoded back to its own rate, so the spectrogram spans the file's real Nyquist and the FFT resolution follows it (bin width = rate/1024). Choose an explicit 8–48 kHz to force a resample instead; that changes playback, the drawing and the exported metadata. Export the resampled audio with Export ▸ Download audio as WAV.
Spectrogram — auto-ranged to the loudest bin on screen, so quiet files are not blank and loud ones are not saturated. The FFT runs in a worker, so zooming and panning stay smooth while it catches up; where a worker cannot be created it falls back to computing inline.
Undo — every span edit is undoable: adding, deleting, retagging, relabelling, dragging, importing, and Clear all. Typing a label undoes as one step, not one keystroke at a time. Each file carries its own history.
Export — JSON, CSV, windows.json (keep/suppress windows for an eval harness), or Audacity
labels, per file or per folder.
Tags — edit the palette in the left rail; it is kept in localStorage and survives reloads.
Importing a span file creates any tags it references. Reset restores the built-in set.
- The
<link>to Google Fonts at the top of the file is optional — it supplies Spectral and IBM Plex when online. Delete that one line for a fully offline file; the fallback stacks take over. localStorageis per-origin, so opening from disk (file://) and through a local server are separate stores for tags, layout and theme.- Writing sidecars into a folder needs the File System Access API (Chromium). Elsewhere a folder loads read-only and you export by download instead.
- Imported spans are trimmed to the audio length, and any starting past the end are dropped — otherwise they would sit off the timeline where they cannot be seen or dragged.