Step 1: open Vowen’s models folder
Quit Vowen completely first (right-click the tray or menu bar icon and choose Quit).
Settings > Developer holds the actions that open Vowen's own folders.
- Windows
- macOS
Press Win + R, enter the following, and press Enter:Open the
models folder inside it. If it does not exist, create a folder named exactly models. The full path looks like:The application-support folder is lowercase
vowen, not Vowen. Both Windows and standard macOS volumes are case-insensitive, so either spelling will find it, but the folder on disk is lowercase.Step 2: download the model you need
Whisper models
Each link downloads a single.bin file. No extraction. The same files work on Windows and macOS.
English-only variants, slightly better if you only dictate in English:
Parakeet models
Parakeet ships as a.tar.gz archive you extract yourself. The archives differ per platform, so take the right one.
- Windows
- macOS
Parakeet Japanese and Parakeet Mandarin are macOS only. There is no Windows build.
Do not rename the extracted folders. Leave each one exactly as the archive produces it, matching the “Extracts to” column, and Vowen will find it.
Nemotron models (macOS only)
Nemotron streams natively with punctuation and capitalisation in the live preview. Apple silicon only, and there is no Windows build.Parakeet streaming bundle (macOS only)
Parakeet V2 and V3 need a second, separate model to produce a live preview while you speak. Vowen normally downloads this silently in the background after any Parakeet batch model finishes, and deletes it when the last Parakeet model is removed, so you will not see it in the model list. On a restricted network that background download fails silently and you get batch transcription with no live preview.
The extracted folder must contain a
320ms subfolder with the bundle’s files inside it. If that subfolder is missing or empty, Vowen treats the bundle as not installed.
The same bundle serves both Parakeet V2 and V3. Nemotron does not need it, since it streams natively. Parakeet Japanese and Mandarin do not use it either.
Speaker diarization bundle (Windows only)
Windows fetches an offline diarization bundle the first time you run a local diarization. macOS does not use this file.
The extracted folder must contain its two model files directly inside it, not nested one level deeper.
Step 3: place the files
Whisper (.bin files)
Move the.bin file directly into the models folder. Do not rename it.
Everything else (.tar.gz archives)
Extract the archive into themodels folder.
- Windows
- macOS
Open PowerShell and run, adjusting the filename to the one you downloaded:7-Zip and WinRAR work too. Extract twice if needed, since
.tar.gz unpacks to .tar first.The result must be a folder with exactly the name in the “Extracts to” column, containing the model files directly inside it, not nested one level deeper.Verifying a download (optional)
A file you download by hand is not checked for you. If you want to be sure it arrived intact, compare its SHA-256 checksum against the list below before extracting.- Windows
- macOS
SHA-256 checksums for every archive
SHA-256 checksums for every archive
.bin files have no published checksum. Check them against the sizes in the table above instead.Step 4: restart Vowen
Start Vowen and open the Speech models page in the left sidebar. The model should now show as downloaded. Select it and try a dictation.Speech models is a top-level sidebar page, not a section inside the Settings modal.
Troubleshooting
Model not detected after restart- Check the filename matches exactly. Browsers append “(1)” if you download the same file twice, and
ggml-small (1).binwill not be detected. - A partially downloaded Whisper file is ignored rather than used. Compare your file against the sizes in the table above and re-download if it is short.
- Files ending in
.partor.partial, and folders ending in.extracting, are the debris of failed in-app downloads. They are safe to delete.
- Open the extracted folder and confirm the model files are directly inside it. If extraction produced a folder inside a folder, move the inner one up a level.
- Check the folder name against the “Extracts to” column. The name must match exactly.
- The streaming bundle is missing. See the Parakeet streaming bundle section above. Batch transcription works fine without it, which is why this fails quietly.
- Confirm
vowen-diarizer-v1exists in the models folder with its two model files directly inside it.
For network administrators
To make in-app downloads work directly, these hosts must be reachable:huggingface.cofor Whisper modelscdn-lfs.huggingface.coandcdn-lfs-us-1.huggingface.co, which Hugging Face redirects downloads toassets.vowen.aifor Parakeet, Nemotron, the streaming bundle and the diarizer bundle
Allowlisting the domains is not enough if the network performs SSL inspection. Certificate replacement is what produces the “self signed certificate in certificate chain” error, so these hosts need to be exempted from SSL inspection, not merely permitted.