Remove Retakes (Cut Repeated Takes from Talking Videos)

About the Remove Retakes (Cut Repeated Takes from Talking Videos)

Recording a talking-head video, a course, a voice-over or a podcast usually means saying a line again until it comes out right — and then scrubbing through the recording to cut the attempts you did not want. This page finds those retakes for you: it transcribes what you said on your own device (or reads the subtitles you already have), finds the sentences you said more than once within a minute or so, and proposes to keep the last take — or the cleanest, with the fewest “um”s, repeats and cut-off endings — and to cut the others, together with the “sorry, again” in between.

Nothing is cut until you have checked it. Every retake is listed with each take’s words and time and a Play button, so you can listen, choose another take to keep, or untick a retake to keep it all. The cuts go in the quietest moment of each pause, with a short cross-fade in the sound, and the result is a new MP4 (or MP3, M4A or WAV) plus the transcript of what is left, as text or SRT subtitles.

The free preview on this page is made of the parts around the cuts — up to 10 seconds of the video with a MySmartCoPilot watermark, or up to 30 seconds of the sound with a short chime — and the first words of the transcript. With a Pro pass, or after unlocking this one result, you get the whole file and the full transcript.

Common uses

• Talking-head videos for YouTube, Reels and Shorts recorded without stopping the camera
• Online course lessons, tutorials and screen-recorded explanations
• Podcast and interview segments where a question or an answer was asked again
• Voice-overs and narrations read from a script or a teleprompter

How retakes are found

• What is said. The speech-recognition model runs in your browser and gives the recording’s sentences with their times; subtitles you bring work the same way, without AI. Long stretches are split at sentence ends, and a sentence that starts again part of the way in (“so today we, um, so today we’re going…”) is split where it starts again.
• Takes of the same line. A sentence is compared with the four before it, as long as only short asides (“sorry”, “one more time”, “let me try that again”) sit between them and it comes within the time window (1 minute unless you change it). Two sentences are takes of one line when they share most of their words (word overlap and spelling similarity, set by Match), when one is the cut-off start of the other (a false start), or — with the sentence model on — when they mean nearly the same and still share some words.
• The take kept. The last take by default, because people usually repeat a line until it is right; or the cleanest one, with the fewest fillers, words said twice and unfinished endings.
• The cut. From the pause before the first take that goes to the pause before the take that stays, at the quietest 10 ms the sound has there, so no word is clipped; the sound is cross-faded over 30 ms at each join.

Languages, accuracy and checking

Speech recognition covers about a hundred languages, Hindi among them, and subtitles you bring can be in any language or script; matching by meaning uses an English model for English and a multilingual model for everything else. Like any speech recognition it can mishear names, accents, music and people talking over each other, and it sometimes leaves out a false start altogether — so treat the list as suggestions: listen to the takes before you save, and untick anything that is not a retake.

Limitations

• Only retakes said as words are found: a take you repeated because of a cough, a noise or a gesture is not, unless the words differ.
• Cuts follow the sentences of the transcript; a slip inside a long sentence that was never restarted is not cut.
• Speech recognition can miss very short false starts and repeated words, which it tends to tidy away.
• A video is encoded again (an H.264 MP4 at high quality, or WebM where the browser cannot write H.264) so the cuts can fall on any frame. Sound-only results keep the original’s channels: WAV at its bit depth (16 or 24-bit), MP3 or M4A at about its bitrate.
• Up to 90 minutes on a computer and 20 minutes on a phone; cut longer recordings into parts first. Phones take noticeably longer than computers.

Privacy

Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot. The video engine downloads from MySmartCoPilot the first time you open a file. To transcribe on your device, a speech-recognition model (about 44, 80 or 252 MB, by the quality you choose) downloads once after you agree; matching takes by meaning downloads a sentence model (about 24 MB for English, 135 MB for other languages) the same way. Your videos, recordings, subtitles and transcripts are never uploaded.