clicketysplit

A browser-based segmentation tool for speech recordings of single-word or pseudoword stimuli.

You give it a stimulus list (real words, pseudowords, whatever your study needs) and a recording. It proposes word boundaries, auto-labels each token from the stimulus list, and gives you a keyboard-driven review UI with fuzzy-match autocomplete for quick corrections. The output is a folder of cleanly-sliced WAV tokens plus a manifest — optionally a Praat TextGrid so the source recording stays usable in Praat.

Why this exists

This replaces the Praat-based workflow used in many labs for a specific kind of session: single-word tokens read in sequence, which may or may not be real words. That workflow means hand-placing every boundary and typing every filename. For a 60-word list at 2 repetitions each, that's 120 slices per speaker per condition, by hand.

• Stimulus-list aware. You declare what the speaker was supposed to say. Autocomplete and auto-labeling work just as well on pseudowords as on real words — there's no dictionary or forced aligner assuming English lexical items, and nothing to transcribe.
• Fast labeling. The tool knows the presentation order (massed aaa bbb ccc, spaced abc abc abc, or random) and how many repetitions to expect, so it prefills the label it expects next. When the speaker skips ahead, adds an extra repetition, or fumbles a take, the prediction adapts instead of dragging every later label out of alignment.
• Praat-compatible, but you only touch Praat if you want to. Optional TextGrid export keeps the source recording useful for follow-up acoustic analysis. The fast path — propose → review → export — never opens Praat.
• Browser UI, local backend. It's a Python package: pip install, run clicketysplit serve, and the segmenter opens in your browser. No Electron, no cloud, no audio leaves your machine.

Install

pip install clicketysplit          # core: Silero VAD, WAV/FLAC/OGG
pip install 'clicketysplit[all]'   # + webrtcvad, denoise, MP3, Praat TextGrid

Requires Python 3.10+. Optional system dependency: ffmpeg (only for MP3/M4A input).

Try it

clicketysplit demo

This boots a throwaway two-condition experiment against bundled audio — one clip of real words, one of pseudowords — so you can walk the whole flow without setting anything up.

Using it on your own recordings

clicketysplit serve

Point the setup wizard at a folder of recordings laid out by speaker and condition:

recordings/
├── f1/
│   ├── real_words/session.wav
│   └── pseudowords/session.wav
└── f2/
    └── ...
stimulus_lists/
├── real_words.txt
└── pseudowords.txt

The wizard discovers speakers and conditions, matches each condition to a stimulus list by name, and lets you set detection parameters. Then:

Review — one token at a time, entirely from the keyboard:

To adjust a slice, click anywhere on the waveform or the spectrogram to move the nearest boundary there, or drag a boundary handle if you want to see it move. Both views are interactive and stay in sync.

Select — pick which repetitions of each word to keep.

Export — write the tokens out.

Each speaker × condition gets a tokens/ directory of WAVs, a token_manifest.json, and a tokens.csv. Sessions autosave and survive a restart or a move of the experiment folder — paths in the config are relative.

How it works

1. Detect. Silero voice-activity detection proposes speech regions, with optional background-noise reduction first. (webrtcvad is available as an alternative via the [webrtc] extra.)
2. Refine. Each boundary is snapped to the nearest energy-envelope crossing, so slices start and end on the actual acoustic edge rather than the VAD frame grid.
3. Classify. Segments are typed by duration: too short is noise, too long is crosstalk, the rest are words. The crosstalk cutoff adapts to the recording's own median word length, so a slow speaker's ordinary words don't get thrown out.
4. Label. For massed recordings, word tokens are clustered on the pauses between them — speakers pause briefly between repetitions of a word and longer when moving to the next one — and each cluster takes the next stimulus in the list.
5. Review and export. You correct what the detector got wrong, choose the takes you want, and export.

Provenance

The detection engine is a port of a tool I wrote and used to cut the stimuli for a word learning experiment — recording each word and pseudoword several times, then slicing out the individual tokens to use as experimental items. The port is verified against the original's output on those recordings: identical segment counts and 0.0 ms boundary differences. The labeling behavior — how the predicted label reacts when a speaker skips ahead, adds a repetition, or produces a token you reject — is ported from the same tool, having been shaped by actually using it for hours.

The bundled demo audio is my own voice, cut from those recordings.

Development

pip install -e '.[dev,all]'
pytest                       # Python tests
cd frontend && npm ci && npm run test && npm run build

The Python and TypeScript labeling implementations share tests/labeling_test_vectors.json; both suites replay it so the two can't drift.

License

MIT — see LICENSE.