Skip to main content
Local speech models are not bundled with Vowen. You download each one from the Speech models page, and Vowen keeps it on your computer until you delete it. Cloud models need no download.

Your first model

Onboarding asks you to download one model before you finish setup. Parakeet Ultra and Parakeet V3 auto-detect 25 European languages. Parakeet V2 is English only. Click Download model to start. The onboarding model step mid-download, with a progress bar on the chosen Parakeet card After the download, the model is optimized for your computer. This can take up to about 30 seconds and needs no action from you. You can change models at any time from Speech models in the sidebar.

Download a model

1

Open Speech models

Click Speech models in the sidebar and scroll to Local. Models are grouped under English Only and Multilingual.
2

Check the size

Each card shows the model name and its download size. Models you don’t have yet show a Download button.
3

Download

Click Download, or click the card itself. The download starts straight away, with no confirmation. A progress bar with a percentage appears on the card. Click the stop button next to it to cancel.
Clicking a card you have not installed starts the download immediately. Whisper Large v3 is 2.9 GB, so check the size first.
4

Start using it

The model is selected when the download finishes and shows the Active tag.
The Local section of Speech models, with English Only and Multilingual model cards, the Active tag on Parakeet Ultra and Download buttons on models not yet installed If a download was interrupted, the card reads Incomplete download with a warning icon. Click it to download again.
On corporate or restricted networks, downloads can fail with a certificate error. See Manual model installation to download models in your browser and install them yourself.

First-run optimization

The first time a Parakeet or Nemotron model runs on your Mac, it is compiled for your chip. This happens once per model. Parakeet Ultra’s first compile takes around 15 to 17 seconds; older Macs take longer. While it runs, the card shows Optimizing model… and a banner reads Optimizing Parakeet model: “First-run optimization for your device. Transcriptions will work once this finishes.”
A dictation you start during optimization is saved, not transcribed. Its History row reads “Transcription pending — model is optimizing”. When optimization finishes, open the row’s menu and choose Retry transcript. See History.
After that, the model stays loaded between dictations, so there is no wait.

The hidden streaming model

On macOS, Parakeet V2, V3, Ultra and Phonon-2 need a separate streaming model for the live preview. It is about 198 MB. Vowen downloads it in the background after the main model finishes, and does not show it in the list.
  • One copy is shared by all four models.
  • It is removed when you delete your last model that uses it.
  • Parakeet Japanese and Parakeet Mandarin do not use it.
If you find a parakeet-eou-streaming folder in your models folder, this is it.

Other background downloads

None of these appear as selectable models.

Switch and delete models

  • Switch: click any installed model. You can also switch from the tray menu under Speech Models.
  • Delete: click the trash icon on an installed model’s card. The active model has no trash icon. To delete it, switch to another model first.

Disk space

Add about 198 MB for the streaming model if you use Parakeet V2, V3, Ultra or Phonon-2 on macOS. Parakeet uses a different build on each platform, which is why its sizes differ. Parakeet Ultra, Phonon-2, Parakeet Japanese, Parakeet Mandarin, Nemotron and Cohere Transcribe are macOS only and do not appear on Windows.

GPU acceleration (Windows)

Local Whisper on Windows can use an NVIDIA GPU.
  • Needs an NVIDIA RTX 2000-series card or newer. AMD and Intel GPUs are not supported.
  • The CUDA module is about 631 MB. You don’t need to install CUDA separately.
1

Find the section

Scroll to the bottom of Speech models. GPU Acceleration appears only when Vowen detects a supported card.
2

Download

Click Download GPU acceleration. Progress is shown in MB.
3

Restart Vowen

The card then reads GPU active · RTX 2000+. To remove it, click Remove GPU acceleration.

Resource efficient mode

Settings › General › Resource efficient mode runs the model as a one-off process for each transcription instead of keeping it loaded. Memory use drops and transcription is slightly slower. The row appears only when it applies: with a Whisper model on macOS or Windows, or with Parakeet on Windows. On Windows with Whisper, GPU acceleration is not available in this mode.

Cloud models

Cloud models use no disk. Open Speech models, click a provider under Cloud, paste your API key and click Save. See Connect a cloud model.