01
Transcript-based cutting worked
The interface reported Remove fillers (3). After the one-click filler action, 6 false-start words were selected manually, producing 9 marked words and 4 cut ranges. The edited timeline showed 27.6 seconds.
INDEPENDENT TEST · PARTIAL RESULTTESTED 2026-07-29
Transcript-based video editing
Can you cut spoken mistakes by deleting words instead of editing a timeline?
In a local build of an exact commit from Rescript's official repository, a 30.77-second self-made English sample marked 3 filler words and 6 false-start words in the transcript, producing 4 cut ranges and a 27.64-second MP4. Project recovery and cached offline re-export worked, but a fresh offline transcription failed and the exported media still requires manual listening.
- Original
- 30.77s
- Words marked
- 9
- Export
- 27.64s
- File size
- 2.237×
1280×720 / 30fps
4 cut ranges
Observed ≤5.1s
Compared with original
INDEPENDENT TEST · PARTIAL RESULTEXACT-COMMIT LOCAL BUILD · 63fa992LIVE GITHUB PAGES NOT TESTED
The same sample: from a 30.77-second transcript to a 27.6-second edit state
The left image shows the original state after Whisper Base transcription. The right shows the interface after 3 fillers were marked in one click and 6 false-start words were selected manually. The pair proves that the edit state and cut ranges changed, not that every word completely disappeared from the audio.
Before: original transcriptORIGINAL STATE
Marked edit state: 4 cutsMARKED EDIT STATE
When the export was transcribed again with the same Whisper Base model, three positions were still recognized as “Ah” and “Actually” still appeared. Because the same model judged both passes, this is not an independent comparison and does not prove degraded audio, failed cuts, or transcription accuracy. It is only a warning to listen to every edited position manually.
- Words marked in UI
- 9
- Export duration
- 27.64s
- Export file size
- 2.237×
4 cut ranges
Original 30.77s
Compared with original
02 / WHAT ACTUALLY WORKED
Separate observed results from what still needs checking
Screenshots, media parameters, and recovery checks show how far the workflow got. They do not replace listening to every edited position.
02
The MP4 decoded, but the file became larger
Export completed within the observed 5.1-second window. The 27.64-second, 1280×720, 30fps H.264/AAC MP4 decoded fully and showed no visible watermark. At 10,338,087 bytes, it was 2.237 times the original file size.
03
The current project recovered and re-exported offline after caching
The project remained after refresh. With the model and project already cached, an offline re-export completed in 2.452 seconds. A fresh offline transcription failed, so this is not a first-run fully-offline result.
Tested-build boundary
This test covered only a local build of official repository commit 63fa992272d69e4fc3cf6a5ce49ca658cbafde89. No media was uploaded to or processed in the public GitHub Pages version. The public GitHub Pages build was not tested. Local-build results must not be presented as results from the live surface.
03 / EXPORTED RESULT
What exported—and what changed in the export
This MP4 decoded fully and showed no visible watermark, but audio was re-encoded as AAC 22050 Hz mono and the file grew to 2.237 times the original size.
04 / MANUAL LISTENING REQUIRED
Same-model retranscription is only a manual-listening warning
When the export was transcribed again with the same Whisper Base model, three positions were still recognized as “Ah” and “Actually” still appeared. Because the same model judged both passes, this is not an independent comparison and does not prove degraded audio, failed cuts, or transcription accuracy. It is only a warning to listen to every edited position manually.
Do not read this image as proof that automatic removal failed—or that it was accurate. Whisper Base judged both passes, so this is only a risk signal that requires manual listening.
05 / IS IT A FIT?
Decide from your task conditions first
This result supports one narrow use case. It does not generalize to every language, duration, or delivery requirement.
Worth a first test
- You have a short, non-sensitive English explainer, podcast, or interview and want to locate mistakes and pauses from a transcript.
- You can use a current Chrome or Edge browser, wait for the first model download, and manually review every edited position after export.
- You accept re-encoding and a potentially larger output, and will keep the original media for comparison.
Not supported by this test
- You need first-run transcription while offline or cannot pre-download the roughly 200 MB / 600 MB model plus supporting models.
- You need a validated path for Chinese, long multi-speaker media, lossless audio, or batch production.
- You cannot manually listen to the output or need automatic filler removal to be final without review.
06 / LIMITS PROMOTION CAN MISS
Do not extend the result beyond these bounds
Only a 30.77-second self-made English sample was tested; Chinese, long videos, multiple speakers, complex background audio, and real meetings were not tested.
A cached offline re-export does not establish first-run offline use or a complete audit of media-related network requests.
No visible watermark was observed in this export only; other formats, settings, and future versions were not tested.
The export used AAC 22050 Hz mono and was 2.237 times the original file size. No independent perceptual, audio-quality, or synchronization measurement was completed.
The README says MIT, but the exact-commit checkout has no LICENSE file; it also uses @ffmpeg/core-mt, whose package metadata says GPL-2.0-or-later. The overall licensing boundary still needs review.
07 / MINIMUM-COST TRY
Run one minimum test with a non-sensitive short clip
Do not begin with meetings, client work, or unpublished material. Use a disposable 20–30-second self-made clip containing one filler and one false start you intend to remove.
- 01
Record the exact commit or live version, browser, model, first-download time, and network state.
- 02
Remove only one filler and one false start, and confirm that matching cut ranges appear.
- 03
Play the export from start to finish and listen closely at every cut. Do not substitute same-model retranscription for manual review.
- 04
Compare duration, resolution, audio parameters, file size, sync, watermark, and recovery after refresh. Stop if any critical result is unacceptable.
The public live build was not covered by this independent test
Open the public Rescript appIf you still choose to try it, use only a disposable, non-sensitive clip you made yourself. Complete the minimum test above before using meetings, client work, or unpublished media.
08 / ORIGINAL SOURCES
Return to the repository, pinned commit, and launch page
Maker claims, BitShovel's local observations, and unresolved items are kept separate. Open the original pages to verify them yourself.
[1]Official Rescript repositoryProject source, README, and public development state.GitHub[2]Exact commit used for this testThe local build was pinned to 63fa992272d69e4fc3cf6a5ce49ca658cbafde89.GitHub[3]README: local processing, model downloads, and browser requirementsMaker statements that are not automatically treated as independent findings.GitHub[4]Rescript on Product HuntPublic launch and discussion surface, not evidence of product results.Product Hunt[5]@ffmpeg/core-mt 0.12.10 package metadataPackage metadata identifies GPL-2.0-or-later, supporting the unresolved licensing boundary.npm