AiViralCut
AiViralCut
Download

Guide · Local captions

Word-level captions that never leave your device

How AiViralCut's captioning works under the hood, what it costs you (one model download), and where it falls short. Everything documented here was captured from the running editor.

What actually happens when you click transcribe

  1. The editor decodes your selected clip's audio and resamples it to 16 kHz mono PCM — the format Whisper expects.
  2. A Web Worker downloads the ONNX model once (Whisper tiny.en ≈ 39 MB, base ≈ 73 MB) and caches it in browser storage.
  3. Transcription runs on your hardware: WebGPU when the browser exposes a GPU adapter, WASM otherwise. Your audio is never uploaded — the model came to you, the audio never leaves.
  4. The engine returns text plus per-word timestamps, which the editor converts to timeline time (respecting clip offsets and speed) and places on track T1 automatically.

Editing the captions it produced

Captions are ordinary clips: select one and the properties panel exposes every word with its start and end time. Rename a word, nudge its timing, delete filler, or add a word the model missed. Style — the kinetic presets — can be switched at any time without re-transcribing.

First-hand captures

Recorded from the production build of this site: a real browser session driving the real editor, with no cuts or retakes. Output files below were produced by the editor's own export engine.

Real session on the production build: text clip creation, all four caption styles with their preview cards, and the local Whisper transcription screen with model and audio-source options.

Supported formats

  • Audio sources: the selected timeline clip, an uploaded audio/video file, or the bundled demo tracks
  • Models: onnx-community/whisper-tiny.en (English, ≈ 39 MB) and onnx-community/whisper-base (multilingual, ≈ 73 MB)
  • Output: word-level caption clips on T1 with per-word timeline timestamps; exportable inside project backups

Known limitations

  • The model downloads on first use — expect a one-time wait proportional to your connection
  • WebGPU requires a Chromium-based browser with GPU access; Safari and older Firefox fall back to WASM, which is slower on long clips
  • Heavy music over speech reduces word accuracy; captions of instrumentals are meaningless by definition
  • Accents, crosstalk and slang cause real errors — use the per-word editor rather than trusting raw output

Ready to try it: open the editor, add a clip with speech, and open AI Tools → Captions. For styling choices, read the kinetic caption styles guide.