Open source

multilingual-subtitles

Subtitles in twenty languages, from one video file, in one command.

A local command-line tool. It extracts the audio, transcribes it with a timestamp on every word, builds a valid SRT, then translates that SRT into every language you name. The timecodes are never retyped, in any language.

Role: Author
Year: 2026
Stack: Node with no npm dependencies, ffmpeg, Mistral, AI Glot CLI
Visit: Source code

Transcribing is the half that's already solved

A dozen tools turn a video into one subtitle file. What stops most channels at one language is the other half: twenty languages is a day of retyping, so nobody does it. That's the half this tool is built for, and it's why the name says multilingual rather than transcription.

Words, not segments

The same Mistral endpoint returns whole segments or individual words, at the same price. A segment can be seven seconds of speech in one block, which is a sentence and not a subtitle: turning it into cues means guessing where inside those seven seconds each phrase was said, and the guess drifts. With word timings the cue builder groups on what reads well, and every cue still starts exactly on the word it shows.

Why an agent shouldn't translate the SRT

For one language, let it. For twenty it's the wrong tool three times over: the model retypes every timecode and eventually mistypes one, the whole file crosses the context window once per language, and you pay generation prices for a mechanical job. AI Glot reads SRT natively, keeps the numbering and the timecodes byte for byte, and costs one credit per three words.

What the rerun costs

Nothing you already paid for. A manifest holds the batch id and status per language, so a language already on disk is skipped and a batch that was created but never downloaded is resumed instead of bought twice. One language failing doesn't stop the others.

One command, and the timecodes are never retyped

The pipeline is five steps, and only the third one is worth having an opinion about.

npm run subs -- run ~/Downloads/clip.mp4 --to fr,de,ja

1Extract the audioffmpeg, mono 16 kHz
2TranscribeEvery word gets a start and an end
3Build the cuesWhere the subtitle quality is decided
4Validate the SRTBefore anything is paid for
5TranslateOne batch per language, four at a time

Each step writes a file and stops. Transcription costs cents and translation costs credits, so keeping them apart means the expensive one never runs just because the cheap one worked.

Step three is lib/cues.mjs, and it is the whole file that makes the difference between a transcript and a subtitle. It breaks on sentence ends and on pauses over 0.6 seconds, holds each cue 1.2 seconds minimum without ever biting into the next one, and caps a cue at two lines of 42 characters and six seconds.

An agent doing itOne language at a time

Retypes every timecode, and eventually mistypes one
The whole file crosses the context window, once per language
Generation prices for a mechanical job
Nothing to resume if it stops halfway

A batch translatorTwenty at a time

Numbering and timecodes kept byte for byte
Four batches in flight, none of it through a model context
One credit per three words
A manifest, so a rerun pays for nothing twice

For one language, letting an agent do it works. The costs on the left are the ones that only show up at twenty.

50languages, ordered by YouTube audience

34credits per language on a 101 word clip

0npm dependencies