Translating hour-long and two-hour videos with AI

What changes when a video passes the hour mark

Nothing about the translation itself changes with length; a sentence in minute 95 is handled exactly like a sentence in minute 2. What changes is everything around it. There is more audio to decode and upload, more chunks to keep consistent, more output to check, and more time during which a closed laptop lid or a dropped connection can interrupt the job. Long-form content also tends to be harder material: lectures with dense terminology, panels with several voices, livestreams with music and game audio.

How long audio is chunked behind the scenes

Speech recognition models handle audio in windows of about 30 seconds, so a two-hour file is processed as roughly 240 separate pieces. In mydubly the browser first decodes the whole soundtrack once to 16 kHz mono audio (24 kHz for a dub), about 230 MB for two hours at 16 kHz, reading the video file in place rather than loading it into memory, which is why very large source files don't crash the tab. It then cuts the audio near every 30-second mark at the quietest 50 ms within the preceding 6 seconds, so sentences are not split mid-word.

The chunking article explains why cutting at pauses beats cutting on the clock.

Splitting recordings longer than 2 hours

mydubly accepts files of up to 2 hours each. For a three-hour conference day or a long livestream, split the file first.

1. Find natural break points: a session change, a coffee break, a pause between questions. Cutting inside a sentence costs you that sentence in both parts.
2. Export each part from your editor, or cut without re-encoding using ffmpeg, for example: ffmpeg -i day1.mp4 -t 01:45:00 -c copy part1.mp4 and then ffmpeg -ss 01:45:00 -i day1.mp4 -c copy part2.mp4.
3. Be aware that cutting without re-encoding snaps to the nearest keyframe, so parts may overlap by a second or two. Check both sides of the cut.
4. Keep each part at least 2 minutes long if you want a voice track, since dubbing has a 2-minute minimum per file.
5. Translate each part into the same target language with the same voice so the parts sound continuous.
6. If you rejoin the parts, shift the second part's subtitle timestamps by the duration of the first part. Free subtitle editors such as Subtitle Edit or Aegisub can offset a whole file at once.

Often the simpler answer is not to rejoin at all. A three-hour event published as separate sessions is easier to watch, easier to review and easier to fix if one part needs re-running. Trimming dead air before splitting also saves money, because you pay per minute of audio; the livestream article covers trimming in more detail.

Keeping terminology consistent across chunks

Because translation happens chunk by chunk, each piece sees around half a minute of context. A product name, a technical term or a person's title can come out one way in the first half hour and another way later. In a short video you'd notice; in a 90-minute lecture you might not.

• Before translating, write down the 10 to 20 terms that matter most: names, products, key concepts, acronyms.
• After translating, search the translated transcript for each term and for its likely alternatives.
• Fix inconsistencies in the subtitle files, which are plain text and easy to search and replace.
• For the voice track, decide whether a variant is wrong or merely different. Two acceptable synonyms are rarely worth a re-run; a wrong term in a key passage may be.

Translating names and technical terms goes deeper on glossaries and transliteration.

Keeping the browser tab alive for the whole job

In mydubly the browser does real work at both ends: it decodes and uploads the audio at the start and assembles the translated video at the end. For a long file, set the device up before you begin. Plug in the laptop, turn off sleep or keep the lid open, and avoid putting the computer into low-power mode. On a phone, keep the screen on and stay in the browser, because mobile browsers commonly pause tabs in the background. Large outputs can be written straight to disk, so make sure there is free space for a file roughly the size of the original video.

Reviewing a long transcript without reading every word

A two-hour translated transcript is a long document. Read it strategically.

1. Search first. Look up every name and key term from your glossary, in both the source and translated transcripts.
2. Scan for anomalies: lines repeated several times in a row, sudden very short or very long segments, or text during stretches you know were music or silence, all signs of a recognition problem.
3. Read the opening five minutes and the closing five minutes in full; viewers judge quality there.
4. Spot-check a minute from every 15 minutes in between, jumping by timestamp.
5. Listen to the dubbed version only at the points flagged in the first four steps.

The guide to proofreading an AI transcript has more search tricks.

Trade-offs of translating long-form content with AI

Long recordings magnify the weaknesses of automated translation. A single voice reads a two-hour panel, which makes it harder to follow who is talking than in a short clip. Inconsistent terminology across chunks is more likely simply because there are more chunks. A recognition problem caused by a bad microphone during one segment affects that whole segment. And the longer the content, the less likely anyone reviews all of it, so the risk of an unnoticed error goes up. For dense, high-stakes long-form material, translated subtitles plus a careful review often serve better than a full voice track.

Long videos in mydubly at a glance

Maximum length

2 hours per file; split anything longer.

Formats

MP4, MOV, WebM, MKV and M4V video; MP3, WAV, M4A, AAC, OGG and FLAC audio.

Price for 90 minutes

4,500 credits ($4.50) with a voice track, or 90 credits (9 cents) for transcripts and subtitles.

What leaves the device

Compressed audio only; the video file stays local.

Retention

Results deleted within 30 minutes of the job finishing; download straight away.

Getting started with a long recording

Trim what you don't need, split at natural breaks if the file runs past two hours, and set your device so it won't sleep. Then open the video translator, choose the target language and let it run. For lecture-specific advice, see lecture translation.