Cut subtitles
straight from the file.

What comes out

Four artefacts,
one transcription.

Ask for as many outputs as you like in a single command. They are all rendered from the same cached transcript, so extra formats cost nothing but the write.

1
00:00:12,040 --> 00:00:15,900
Nobody tells you that the hard part
of subtitling isn't the words.

2
00:00:16,400 --> 00:00:18,880
It's deciding where to cut them.

WEBVTT

00:00:12.040 --> 00:00:15.900
Nobody tells you that the hard part
of subtitling isn't the words.

00:00:16.400 --> 00:00:18.880
It's deciding where to cut them.

Nobody tells you that the hard part of subtitling
isn't the words. It's deciding where to cut them.
Every provider hands back a wall of tokens with
timestamps attached, and the shaping is left as an
exercise for the reader.

[
  {
    "index": 0,
    "text": "nobody",
    "punctuated": "Nobody",
    "start": 12.04,
    "end": 12.44,
    "confidence": 0.99873,
    "speaker": 0,
    "channel": 0,
    "language": "en-US"
  },
  { "index": 1, "text": "tells", … }
]

The pipeline

Five stages.
Each one an artefact.

Every stage writes a cached artefact keyed by the source media hash, the step options, the pipeline version and — for audio — the ffmpeg build. Stop at any one of them, or run the lot.

Cue shaping

The hard part
isn't the words.

Transcription gives you tokens and timestamps. Deciding where a cue starts, where a line breaks and how long it stays on screen is a different problem — and subkit ships two answers to it. Same audio, same transcript, below:

1
00:00:12,040 --> 00:00:14,180
[speaker 0]
Nobody tells you that the hard part of

2
00:00:14,180 --> 00:00:17,640
subtitling isn't the words. It's deciding where to

3
00:00:17,640 --> 00:00:18,880
cut them.

1
00:00:12,040 --> 00:00:15,900
Nobody tells you that the hard part
of subtitling isn't the words.

2
00:00:16,400 --> 00:00:18,880
It's deciding where to cut them.

deepgram is a faithful port of deepgram-go-captions — fixed-size chunks, timing taken straight from the first and last word. netflix follows the Timed Text Style Guide: it never spans a speaker change or a long silence, packs whole sentences while they fit, splits oversized ones at the best linguistic break, and favours bottom-heavy line breaks. Read the full comparison →

Caching

subkit subtitle movie.mp4 --rerun transcribe  # one stage
subkit subtitle movie.mp4 --rerun audio,cues  # several
subkit subtitle movie.mp4 --refresh            # rebuild all
subkit subtitle movie.mp4 --no-cache           # touch nothing

subkit cache path                       # where it lives
subkit cache list                       # what's in it
subkit cache clean                      # empty it
subkit --cache-dir ./.subkit-cache subtitle movie.mp4

Distribution

Install it

A single static binary, built for amd64 and arm64 on all three platforms. No runtime to install beyond ffmpeg.

winget install AndreSantos.Subkit

subkit_darwin_arm64.dmg

curl -sSL …/subkit_linux_amd64.tar.gz | tar xz

go install github.com/andrerfcsantos/subkit-codex/cmd/subkit@latest

Prerequisites: ffmpeg on your PATH and a DEEPGRAM_API_KEY environment variable. Per-platform notes, verification and shell setup are in the install guide.