Generate locally. Review every word.

Studio Beta in 1.2.0

Generate from a clip

1. 1

   Choose the source

   Select the video or audio clip that contains speech, then open Captions in the library.

2. 2

   Set the spoken language

   English is the default. Select the language actually spoken in the source, including Polish when appropriate.

3. 3

   Download / verify model

   The first use requests access to known delivery hosts and downloads the pinned speech model, roughly 80 MB. Model data is stored separately from projects.

4. 4

   Choose the processor

   Auto (GPU + CPU) can accelerate audio analysis on WebGPU; speech decoding stays on CPU / WASM. CPU / WASM can also be selected directly.

5. 5

   Generate from selected clip

   Follow preparation, model-loading and transcription stages, source part, completed audio, actual decoder activity and elapsed/last-update times. Cancel keeps the project.

6. 6

   Review the result

   Generation creates a caption layer as one Undo operation. Correct words and timing before export. Downloaded runtime code is not used: executable libraries are bundled locally.

Create and exchange subtitles

Add manual cues or import SRT/VTT. Captions live on their own timeline layers. Edit cue content, start/end, word times and style; split, merge or shift cues. Export SRT/VTT separately when you need subtitle data rather than burned-in video text.

Review captions as a transcript

1. 1

   Open Review captions

   Search all caption text, replace matching text and follow playback in the review panel.

2. 2

   Seek or mark a passage

   Click a word to seek. Selecting an utterance sets In/Out; correcting text does not cut video.

3. 3

   Reflow captions

   Default grouping uses up to four words and two lines, with punctuation boundaries. Reflow retains your corrected words and existing word timestamps where possible.

4. 4

   Keep changes safe

   Inserted words receive estimated timing. Retry does not automatically replace manually corrected generated captions. Undo returns to the previous version.

Restore word highlighting

Text edits try to preserve matching saved word times. New or unaligned words receive estimated timing. If an older edited cue loses highlighting, select it and choose Restore word highlighting in the Captions inspector. Adjust individual word times if the estimate needs correction.

Style within safe areas

Choose fonts, size, outline, shadow, background and active-word highlighting. For new 9:16 projects, guides default to top 10%, bottom 22%, left 6% and right 16%. Adjust them for your target platform; they are guides, not a platform guarantee.

Guides support snapping and out-of-margin text warnings. They never appear in exported frames and never reposition existing layers by themselves. A full template can retain caption style, language, segmentation and safe-area settings.

Offline use and model cleanup

After the model has downloaded and verified, recognition processes audio locally and can run offline. Remove downloaded model deletes that optional data without deleting projects or already generated captions. Project/template files do not include speech-model weights or executable code.