fablecut_transcribe
Transcribe speech from timeline clips into timed transcripts to read interviews, find phrases, and cut them from the edit.
Instructions
What is said, word by word, with times — and text-based editing. Transcribes the clips' speech with Whisper in the editor (an open tab, else headless Chrome / Edge; the speech model downloads once, the audio never leaves the machine), keeps one transcript per file (analysis/transcripts/), and answers with the timeline's words as timed lines (timeline seconds), so you can read an interview, pick the good takes, or find where something is said. find:"phrase" (or a list) reports every place it is said — case, punctuation and spacing ignored, whole words only — with the exact span; cut:true then extracts each of them on the targeted tracks (or tracks), the words and the pause after them, last first, in one revision: later material moves up, linked stems stay in sync, locked clips stay put (lock the music bed first). A version is saved before a cut. Files already transcribed aren't done again unless again:true. Default clips: every video / audio clip with sound on the open timeline. mediaIds reads whole files instead (file seconds; no cut). The same as the editor's Transcript tab.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| to | No | Only words before this time | |
| cut | No | Cut every match of `find` out (default false: only report) | |
| find | No | A word or phrase (or several) to find; matches are reported with their timeline spans | |
| from | No | Only words from this time on (timeline seconds; file seconds with mediaIds) | |
| again | No | Transcribe again even when a transcript exists (another model or language) | |
| force | No | cut: also cut through locked clips (only when the user asked) | |
| model | No | Whisper model: base (default, 77 MB, quick) or small (250 MB, more accurate, slower) | |
| where | No | Where to transcribe: auto (default): an open editor tab, else headless · tab · headless | |
| tracks | No | cut: the lanes to cut on (default: the targeted tracks) | |
| clipIds | No | Clips to transcribe and read (a video brings its linked stems). Default: every clip with sound on the timeline | |
| timeout | No | Seconds to wait for the transcription (default 1800) | |
| language | No | The spoken language as a Whisper code (en, es, hi, ja, …) or auto (default: detected) | |
| mediaIds | No | Instead of clips: whole files, read in file seconds (find works, cut doesn't) |