Skip to main content

Transcribe CLI Tutorial

The Knowii Voice AI desktop app is built around your microphone: you speak, it types. But a lot of the audio worth reading was never spoken into your mic. Meeting recordings. Interviews. Podcast episodes. That conference talk you saved two years ago and never rewatched.

transcribe is the tool for those. It's a command-line program that ships with Knowii Voice AI, runs entirely on your machine, and turns media files into subtitles or plain text.

In this tutorial, you will:

  • Check that transcribe is available and download your first model
  • Subtitle a single video file
  • Get a plain-text transcript you can paste into your notes
  • Transcribe a whole folder in one command
  • Handle a recording in another language

Nothing here leaves your computer. No upload, no account, no per-minute billing.

Prerequisites​

  • Knowii Voice AI installed (see the Getting Started Tutorial)
  • A terminal: Command Prompt or PowerShell on Windows, Terminal on macOS and Linux
  • One audio or video file to practice on
  • Enough disk space for a model (the one used below is about 3 GB; smaller ones exist)

Step 1: Check that transcribe is there​

transcribe is installed alongside the main Knowii Voice AI executable. Open your terminal and run:

transcribe --help

You should see the list of subcommands. If instead you get "command not found", your shell doesn't know where to look yet. Two ways to fix that:

Call it by its full path:

PlatformCommand
Windows"%LOCALAPPDATA%\Programs\knowii-voice-ai\transcribe.exe" --help
macOS"/Applications/Knowii Voice AI.app/Contents/MacOS/transcribe" --help
LinuxThe folder holding the knowii-voice-ai binary from your package

Or make it permanent by adding that folder to your PATH. On macOS, an alias in your ~/.zshrc is quicker:

alias transcribe="/Applications/Knowii\ Voice\ AI.app/Contents/MacOS/transcribe"

The rest of this tutorial assumes plain transcribe works.

Step 2: Get a model​

Transcription needs a model, and models are downloaded once and reused forever. See what is on offer:

transcribe models list
Models directory: /home/you/.local/share/knowii-voice-ai/models
ID                         ENGINE        SIZE  DL   NAME
parakeet-tdt-0.6b-v3       Parakeet      785M    -  Parakeet V3
whisper-tiny               Whisper        74M    -  Whisper - Tiny
whisper-large-v3           Whisper      2952M    -  Whisper - Large V3
...

A ✓ in the DL column means you already have it. If you have used the app before, one of these is probably already downloaded, and you can skip ahead.

For subtitles, use Whisper. It produces timestamped segments, which is exactly what a subtitle file is:

transcribe models download whisper-large-v3

That model is the most accurate, and the biggest. On a slower machine, whisper-medium or even whisper-small will do fine and finish sooner. Start with what your disk and patience allow. You can always download another one later.

The app and the CLI share models

Models you download here appear in the desktop app, and models you downloaded in the app work here. There's one folder, one catalog, no duplication.

Step 3: Subtitle your first video​

Point transcribe at a file and name the model:

transcribe file talk.mp4 --model whisper-large-v3

Wait for it to finish, then look next to talk.mp4. There's now a talk.srt.

Open the video in VLC, mpv, or any player that reads subtitles, and the text appears in sync with the audio. Upload the pair to YouTube and the captions come along.

Notice what you didn't have to do: no format conversion, no ffmpeg, no extracting the audio track first. transcribe reads mp4, mkv, mov, m4a, mp3, wav, ogg, opus, flac, and more, directly.

Step 4: Get plain text instead​

Subtitles are great for video. For notes, you want prose without the timestamps. That's the --format flag:

transcribe file interview.m4a --model whisper-large-v3 --format txt

This writes interview.txt. To skip the file entirely and print straight to your terminal, send the output to -:

transcribe file interview.m4a --model whisper-large-v3 --format txt --output -

Which means you can pipe it anywhere:

# Straight into your clipboard on Linux
transcribe file interview.m4a --model whisper-large-v3 --format txt --output - | wl-copy

# Or into a note in your vault
transcribe file interview.m4a --model whisper-large-v3 --format txt --output - > ~/notes/interview.md

This is where a CLI earns its keep. The transcript becomes text in a pipeline, and every tool you already use works on it.

Step 5: Do a whole folder at once​

One file is a demo. A backlog is the real use case.

transcribe file *.mp4 --model whisper-large-v3 --format vtt --output ./subs/

Every .mp4 in the current folder gets transcribed, and the results land in ./subs/. Start it, go make coffee, come back to a folder full of subtitles.

Fifty meeting recordings you never had time to review become fifty searchable text files while you're doing something else.

Step 6: Work in another language​

Whisper detects the language on its own, so a French recording usually just works:

transcribe file entretien.mp3 --model whisper-large-v3

If detection guesses wrong, or if the recording opens with silence and gives it nothing to go on, say it outright:

transcribe file entretien.mp3 --model whisper-large-v3 --language fr

And when you want the English version of a non-English recording, ask for a translation:

transcribe file entretien.mp3 --model whisper-large-v3 --translate

You get English subtitles from French audio, in one pass.

Feed it your vocabulary

Recordings full of product names, acronyms, or technical terms transcribe better if you tell the model what to expect:

transcribe file standup.m4a --model whisper-large-v3 --initial-prompt "Kubernetes, Grafana, Prometheus, OTEL"

The prompt biases the model's vocabulary. Nothing else costs so little for so much accuracy.

Step 7: Make it faster​

Whisper is accurate and not especially quick. When you care more about speed, switch engines:

transcribe file clip.mkv --model parakeet-tdt-0.6b-v3 --int8

Parakeet is a much faster ONNX engine, and --int8 runs it quantized, faster still. Moonshine is faster again for short English clips.

The trade-off is real: Whisper gives you the best timestamps and the best accuracy on hard audio, which is why it stays the default for subtitles. Use Parakeet or Moonshine when you want a rough transcript of a long recording and you want it now.

What you can build from here​

Once transcription is one command, it slots into everything else:

  • A shell loop that transcribes every new recording in a watched folder
  • A cron job that processes yesterday's meetings overnight
  • A note-taking pipeline that drops transcripts straight into your vault, ready to be searched and linked
  • A subtitle step in your video editing workflow, before anything is published

Start simple. Transcribe one file, see the output, then automate the part you found yourself repeating. The tool rewards that order.

Next Steps​

Questions? Email support@knowii.net or join the Knowii Community.