Turns English speech in your video into timed subtitles on your own computer, with Bark 3.0, the speech model from Conway Research. It runs on the CPU, needs no graphics card and no account, and nothing leaves your machine.

What it does

Offline after one download

The 163 MB model downloads once, the first time you transcribe, and is checked against a pinned fingerprint. From then on, no internet is needed.

No graphics card

It runs on the CPU, many times faster than the audio plays: a ten-minute video takes under a minute on a desktop computer.

Punctuation and capitals

Bark 3.0 writes sentences, not a stream of words: commas, full stops, capitals and numbers come out as you would type them.

One subtitle per sentence

Every word has its own time, so each subtitle is a whole sentence, as with AssemblyAI. Turn off One cue per sentence and long sentences are split to stay easy to read.

Models

There are four Bark 3.0 models. Bark 3.0 is the one used unless you choose another under Model; the others download only if you pick them. None of them is inside the add-on: the one you use downloads into Subtitld's shared model cache the first time you transcribe, so updating the add-on doesn't download it again.

Model	Download	Good for
Bark 3.0 (0.6B, verbatim) Default bark3	163 MB	Everyday English speech, word for word
Bark 3.0 Plus (0.6B, verbatim, more accurate) bark3-plus	180 MB	A little more accurate, word for word
Bark 3.0 Clean (0.6B, drops fillers and false starts) bark3-clean	179 MB	Clean subtitles: leaves out um, uh and false starts
Bark 3.0 Pro (0.6B, full precision) bark3-pro	1.3 GB	The most accurate, for the hardest audio

After the download, the add-on prepares the model for your computer, once, in about half a minute. Prepared, each model takes about 650 MB of disk and needs about 1.2 GB of memory while it transcribes, and all four run at the same speed. Clean leaves out fillers like "um" and "uh" and the false starts of a sentence, which is usually what a subtitle should show; choose Bark 3.0 or Plus for a word-for-word transcript.

Long videos are transcribed about 30 seconds at a time, so the progress bar keeps moving and Cancel takes effect within seconds.

Options

In Settings › Add-ons › Bark 3.0 (offline, English) › Configure.

Model

Default: bark3

Which Bark 3.0 model transcribes. Each one downloads the first time you pick it; the others are never downloaded.

CPU threads

Default: 0

How many CPU cores Bark 3.0 uses. 0 picks half of them, at most 8. A new value applies the next time the model loads.

One cue per sentence

Default: on

On: one subtitle per sentence, however long, as in AssemblyAI. Off: a long sentence is split into subtitles of about two lines or 7 seconds at most.

Languages

English only. Leaving the language empty, or on detect, still transcribes English. A project set to another language is refused: use Whisper (offline) for those.

en

Platforms

System	Download	Tested
Linux x86-64	57.9 MB	On real machines and in CI
macOS (Apple silicon)	39.8 MB	In CI
Windows x86-64	34.7 MB	In CI

Linux x86-64, macOS on Apple silicon and Windows x86-64. ffmpeg comes with Subtitld.

Troubleshooting

unsupported_languageBark 3.0 knows only English. For other languages, use Whisper (offline).
networkThe model couldn't be downloaded. Check your internet connection and try again: an interrupted download resumes where it stopped.
model_missingThe prepared model couldn't be loaded. The message names its folder: delete that folder, and the next transcription prepares it again.
bad_paramsSubtitld couldn't hand the audio over, or the model setting is unknown. The message says why.
oomOut of memory. Close other programs and try again.
disk_fullNot enough disk space to prepare the model or decode the audio. Bark 3.0 needs about 900 MB, Pro about 2 GB while it is prepared.
internalSomething failed inside the add-on, often ffmpeg decoding the audio. The message has the details.

License

Add-on

Apache-2.0

Engine

MIT

ONNX Runtime, by Microsoft

Model

CC-BY-4.0

Bark 3.0 weights, by Conway Research, from NVIDIA Parakeet

Fine for commercial work. The add-on is Apache-2.0 and ONNX Runtime is MIT. The Bark 3.0 weights are CC-BY-4.0, from Conway Research, made from NVIDIA's Parakeet model; the license asks for credit wherever the weights themselves are shared, and the add-on's NOTICE file carries it. The subtitles you make are yours.

Bark 3.0 (offline, English)