Parakeet & Nemotron offline, 33 languages

What it does

Using it

To transcribe a file, open the Import tab, pick Parakeet & Nemotron as the Transcription engine, set the Subtitle Language and press Start transcription. Subtitles appear on the timeline while the rest is still being transcribed: the audio goes a minute at a time (half a minute with Nemotron), so the progress bar keeps moving and Cancel takes effect within seconds.

For live captions, open the Audio tab and pick it as the Transcription engine under Recording.

Models

There are three models, and none of them is inside the add-on. Each downloads into Subtitld's shared model cache the first time it is needed, from NVIDIA's own pages on Hugging Face, pinned to one revision and checked against its fingerprint. An interrupted download resumes where it stopped. Updating or removing the add-on keeps them.

The add-on's Model setting decides which model transcribes a file. On Automatic, the default, Parakeet takes the 25 languages it knows and Nemotron the other eight. You can also choose Nemotron for a language both know; Parakeet is still the better choice there.

Parakeet times every word as it hears it. Nemotron times a word when it writes it, near the word's end, so the add-on moves each word back to where it is spoken. On a test film, 76% of Nemotron's subtitles start within half a second of the film's own subtitles, against 85% for Parakeet. Over music and sound effects Nemotron often writes no punctuation, so its subtitles also end at a pause of 0.8 seconds.

Measured on a laptop processor (Intel i7-13700T):

Where models are kept

All three models take about 1.6 GB. Delete the folder to free the space; each model downloads again when next needed.

Live captions

While you record, Nemotron hears a little ahead of each word (560 ms by default; see Live caption delay) and writes words as you speak. Subtitld shows them until the sentence is finished, and then the sentence fills its subtitle. A sentence ends after 0.8 seconds without a new word.

The first time, the 742 MB Nemotron model downloads, and the download's progress shows where the words will be. Live captions come out in lower case and without punctuation more often than files do: for clean text, transcribe the recording afterwards with Parakeet. Speaker labels are for files only.

Options

In Settings › Add-ons › Parakeet & Nemotron (offline, 33 languages) › Configure.

Languages

33 languages. Parakeet knows 25: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. Nemotron knows 28: all of those except Greek, Latvian, Lithuanian, Maltese and Slovenian, and also Arabic, Chinese, Hindi, Japanese, Korean, Norwegian, Turkish and Vietnamese. Live captions cover Nemotron's 28.

Leave the language empty and Parakeet works out which of its 25 is spoken. Nemotron is told the language, and for English, Spanish, French and Portuguese also the variety: en-gb, es-es, fr-ca and pt-br get British English, Spanish from Spain, Canadian French and Brazilian Portuguese. Without a region it expects American English, American Spanish, French from France and European Portuguese. A language neither model knows gets a message listing the ones they do.

• ar
• bg
• cs
• da
• de
• el
• en
• es
• et
• fi
• fr
• hi
• hr
• hu
• it
• ja
• ko
• lt
• lv
• mt
• nb
• nl
• pl
• pt
• ro
• ru
• sk
• sl
• sv
• tr
• uk
• vi
• zh

Platforms

Linux x86-64 (any distribution from 2018 on), macOS on Apple Silicon and Windows x86-64. On x86-64 the processor needs AVX2: Intel from 2013 and AMD from 2015 on. ffmpeg comes with Subtitld.

On a Mac the add-on tries the graphics processor (Metal) once, in a separate process, and uses the processor if that fails. The verdict is kept in the models folder, in metal-check.json. To choose yourself, set PARAKEET_NEMOTRON_DEVICE to cpu or gpu before you start Subtitld.

Each release is built on all three systems, and on each one CI installs the add-on as Subtitld does, downloads the models and transcribes a file with each model, with speaker labels, and live.

Troubleshooting

• unsupported_languageThe model can't transcribe the project's language; the message lists the ones it can. Set Model to Automatic, or pick another engine. Greek, Latvian, Lithuanian, Maltese and Slovenian have no live captions: transcribe the file afterwards.
• network_unavailableA model couldn't be downloaded, or a downloaded file failed its check and was discarded. Check your connection and try again: the download resumes where it stopped.
• disk_fullNot enough disk space for the models or the decoded audio. All three models take about 1.6 GB.
• oomOut of memory. Close other programs: a long film needs about 1.3 GB.
• bad_paramsSubtitld couldn't hand the audio over; the message says why.
• internalSomething failed inside the add-on, often ffmpeg decoding the audio or a model failing to load. The message has the details. If it says the processor lacks AVX2, this computer can't run the add-on.

License

Fine for commercial work. The add-on and NVIDIA's NeMo-Speech.cpp are Apache-2.0. Parakeet is CC-BY-4.0, which asks for credit to NVIDIA wherever the weights themselves are shared. The two Nemotron models are under OpenMDW-1.1, a permissive license for models that puts no limits on what you make with them. The models download from NVIDIA; the add-on doesn't ship them. The subtitles you make are yours.