Local Whisper models
In Local mode, Abe runs OpenAI’s open-source Whisper speech recognition model on your own computer. Nothing is uploaded, and you don’t need an API key or an account.
How downloading works
Section titled “How downloading works”- Abe offers four models: Tiny, Base (the default), Small and Large v3 Turbo.
- A model is downloaded the first time you use it, with a progress bar in Abe. After that it works fully offline.
- Models are stored in Abe’s app data folder on your computer, not in your Documents or Downloads.
- Abe checks each download’s integrity before using it, so a broken or tampered file is never loaded. If a download fails, just try again.
The models
Section titled “The models”| Model | Download size | Speed | Accuracy | Good for |
|---|---|---|---|---|
| Tiny | ~75 MB | Fastest | Basic | Older machines, quick notes, short commands |
| Base (default) | ~150 MB | Very fast | Fair | Everyday dictation on modest hardware |
| Small | ~470 MB | Fast | Good | A strong default for most computers |
| Large v3 Turbo | ~1.6 GB | Slower on CPU, fast on GPU / Apple Silicon | Best | The best accuracy, especially with accents, names and technical terms |
Sizes are approximate. Bigger models are more accurate and handle accents, background noise and unusual words better, but they need more memory and take longer to process each recording.
Which one should I pick?
Section titled “Which one should I pick?”- Not sure? Keep the default, Base, and move up a size if you want better accuracy.
- Apple Silicon Mac (M1 or later): Large v3 Turbo usually feels quick thanks to GPU acceleration. Start there.
- Windows PC with a recent dedicated GPU: Large v3 Turbo, if GPU acceleration is available in your build.
- Recent laptop, CPU only: Small. Try Large v3 Turbo and see if the delay bothers you.
- Older or low-power machine: Base or Tiny, or consider Cloud mode.
You can change models at any time in Settings → Transcription. Each model is downloaded the first time you select it.
Speed and hardware notes
Section titled “Speed and hardware notes”- Processing happens after you release the hotkey, so the wait grows with the length of what you said. Short phrases are near-instant on most hardware with the smaller models.
- Whisper benefits greatly from a GPU. On Macs, Apple Silicon’s GPU is used automatically. On Windows, GPU support depends on how Abe was built; CPU always works.
- Expect memory use roughly in line with the model size, plus some working memory. Large v3 Turbo is comfortable with 8 GB of RAM or more.
- Transcribing on battery uses more power with larger models.
Languages
Section titled “Languages”Whisper models are multilingual. Set your language in Settings for the best results, or leave it on auto-detect if you switch between languages.
Voice Transform in Local mode
Section titled “Voice Transform in Local mode”Local mode covers dictation. Voice Transform needs a language model, so it always uses a cloud or custom provider, even when you transcribe locally. Use a Groq key (recommended), an OpenAI key, or Ollama on your own computer to keep everything offline.
Privacy
Section titled “Privacy”In Local mode your audio is processed in memory on your computer and is never sent anywhere. The only network request is the model download itself.