What "sentence level" refers to

An AI dub passes through four stages: speech recognition, translation, voice generation and timing. The unit that moves through those stages matters. Speech recognition produces lines of text with start and end times, but those lines follow pauses, not grammar: a speaker who hesitates mid-sentence produces two lines, and a fast speaker can run two sentences into one.
Sentence-level dubbing regroups those lines into sentences before anything else happens. Each sentence then travels through translation, voice generation and timing as one unit. The alternatives are to dub each recognized line as it comes, which often means translating half a sentence, or to translate long passages at once, which loses the link between the new voice and the moment on screen.
Why fragments translate badly

Languages arrange a sentence differently. German often puts the main verb at the end of a clause; Japanese, Korean, Hindi and Turkish put the verb last in general. If the first half of an English sentence is translated on its own, the translation engine has to finish a sentence that isn't finished, or leave a fragment that a native speaker would never say.
Voice models have a similar problem. A lone phrase like "and then" or "so basically" gives the model no sentence to shape, and very short clips are where synthetic voices most often stumble, repeat or sound clipped. The broader effect of phrasing on synthetic speech is covered in prosody in text to speech.
Why whole paragraphs are no better

Going the other way, translating and voicing a whole paragraph at once, gives the translation more context but breaks the timing. A 40-second passage voiced as one clip has only one anchor: its start. If the translation runs longer, which is common (see text expansion in translation), every word after the first drifts further behind the picture, and an on-screen action described at the end of the passage arrives late. Speaker changes inside a passage also need separate clips, because each person is voiced differently.
A sentence is the compromise: long enough to translate and phrase naturally, short enough to keep a timing anchor every few seconds.
How a sentence is timed

Once a sentence has been translated and voiced, the clip has a length of its own, which rarely matches the original. Timing fit is the set of rules that reconcile the two.
- The clip starts where the original sentence started.
- If it is too long for the gap before the next sentence, it can start slightly early or run slightly into the next sentence's time, when that time is free.
- Only then is it sped up, by a small amount that keeps the pitch natural.
- If the clip is shorter than the original, its pauses are widened a little and the voice slowed slightly, so it ends with the speaker rather than racing ahead and leaving silence.
Because every sentence is re-anchored to its own start time, a small overrun in one sentence never accumulates into drift across the video.
A presenter says "So once you've exported the file" (00:12.4 to 00:14.1), pauses for a fifth of a second, then says "open the settings panel and choose Audio." (00:14.3 to 00:16.9). The first line has no closing punctuation and the pause is short, so the two lines are joined into one sentence from 00:12.4 to 00:16.9. It is translated whole into Spanish as "Una vez que hayas exportado el archivo, abre el panel de configuración y elige Audio.", voiced as one clip in the presenter's copied voice and placed at 00:12.4. The Spanish runs a little longer than the English, so the clip may start a fraction early, use the free pause after it, or speed up slightly to finish near 00:16.9.
How mydubly dubs sentence by sentence
mydubly's AI dubbing works this way. After Whisper recognizes the speech, recognized lines are joined into one sentence when the first doesn't end with sentence punctuation, the same person keeps talking, and the pause between them is under 0.7 seconds. A joined sentence is capped at 12 seconds and 260 characters, so a speaker who never pauses is still split into manageable units. Very short fragments of one or two words are joined to a neighbouring line where possible, because a lone word voices poorly.
Each sentence is translated as a whole and spoken in a copy of its speaker's own voice, in any of 19 languages. A line may start up to 0.3 seconds early or run up to 0.6 seconds late, and the tempo stays at or below 1.15× by default, with a little extra only when a stretch can't otherwise fit. Each line's loudness follows the original line, and the original music and effects stay underneath, turned down while the new voice speaks.
Long videos are processed in chunks of roughly 30 seconds cut at quiet moments. When a cut still falls inside a sentence, the end of that sentence is handed to the chunk before it, so the sentence is translated and spoken in one piece. The translating videos with multiple speakers article covers how lines are assigned to each person.
A full translation costs 120 credits per minute with a 2-minute minimum, so a 6-minute explainer costs 300 credits ($0.30) and includes the dubbed video, the translated audio, subtitles and transcripts in both languages.
Sponsored · mydubly
AI video dubbing deal
Dub any video for just 12¢ a minute
Was 240 credits/min — now only 120. Cloned voices, timed to your picture, at half the list price.
Sentence-level editing: limits in mydubly
In dubbing editors, "sentence-level editing" usually means a timeline of the translated sentences where you can rewrite one sentence, change its timing or voice, and regenerate only that clip. It is the fastest way to fix a single mistranslated line without paying for, or waiting for, a whole new dub.
mydubly has no such editor. You can't change one sentence and regenerate it; spoken-audio fixes need a new run of the file at the same per-minute rate. What you can edit directly are the text outputs: the subtitle files and transcripts are plain text, so wording and timing can be corrected in any text or subtitle editor, as described in how to edit an SRT file.
Two habits keep re-runs rare:
- Fix problems at the source. Mispronounced names, mumbled words and crosstalk cause most bad sentences; a cleaner recording or a re-recorded line fixes them for every language at once.
- Review before publishing. Read the two transcripts side by side, then listen at the timestamps of names, numbers and on-screen actions. The AI dubbing quality checklist gives an order for those checks.
Where sentence-level dubbing falls short
- Context beyond the sentence. A sentence translated on its own can lose what earlier sentences established, such as who "she" is or whether the speaker uses formal address; context in machine translation explains why.
- Punctuation from speech recognition decides where sentences end. Run-on speech with little punctuation produces long sentences that hit the length cap and are split at a less natural point.
- Very fast speakers leave little free time around each sentence, so longer translations are sped up more and can sound brisk.
- Overlapping speakers produce lines that belong to two people, which no sentence grouping can untangle cleanly.
- Timing follows sentence starts, not lip movements. Without lip-sync, close-up shots still show mouths that don't match.
Next step
To hear sentence-level timing on your own material, dub a short clip with mydubly's AI dubbing: a 2-minute clip costs 100 credits, the same as the free credits you get on sign-up. Listen especially to the places where the speaker pauses mid-sentence and to on-screen actions, since that is where sentence grouping and timing show most.
Sponsored · mydubly
240 free credits + 50% off
Try free, then buy minutes at half price
New accounts get 240 credits (enough for a 2-minute dub). When you top up, you keep these sale rates.
Hear your video in another language
Pick a target language; mydubly speaks each person's lines in a copy of their own voice, fits them to the original timing and rebuilds the video in your browser.
Frequently asked questions
What is sentence-level dubbing?
It is a way of building an AI dub one sentence at a time. Recognized speech is grouped into sentences, and each sentence is translated as a whole, voiced as one clip and placed at the moment the original sentence started. It balances natural translation with timing that stays close to the picture.
Is sentence-level dubbing the same as lip-sync?
No. Sentence-level timing places each translated sentence where the original sentence started, so the voice follows the speaker's rhythm. Lip-sync changes the video so mouths match the new language. mydubly does sentence-level timing and leaves the picture unchanged.
Can I edit a single sentence in a mydubly dub?
Not in the spoken audio. mydubly has no sentence editor, so a fix to the voice track needs a new run of the file. The subtitle files and transcripts are plain text, so you can correct their wording and timing yourself in any text or subtitle editor.
Why does my dub sometimes start a word before the speaker?
Translations are often longer than the original. To avoid speeding a sentence up too much, the clip can start slightly early when the moment before it is free. In mydubly this is at most 0.3 seconds, which most viewers don't notice.
Does sentence-level dubbing work for languages with very different word order?
Yes, and that is one of its main advantages. Translating whole sentences lets the translation move the verb or reorder clauses as the target language requires, which fragment-by-fragment translation can't do. Check long or run-on sentences in the translated transcript, since they are split at a length cap.
Sponsored · mydubly
Transcripts & subtitles deal