Models — local vs cloud
Choosing a transcription model. Privacy, speed, accuracy, language support.
Whiskers supports local models (run on your Mac, no internet, no cost) and cloud models (faster on long files, no disk usage, but audio leaves your device).
Local models
| Model | Languages | Size | Best for |
|---|---|---|---|
| Parakeet V2 | English | ~400 MB | Fastest local English |
| Parakeet V3 | 25 | ~600 MB | Multilingual, fast, accurate |
| Whisper Tiny | 99+ | ~75 MB | Lowest disk; basic accuracy |
| Whisper Base | 99+ | ~150 MB | Light footprint |
| Whisper Small | 99+ | ~500 MB | Balanced |
| Whisper Medium | 99+ | ~1.5 GB | High accuracy |
| Whisper Large v3 | 99+ | ~3 GB | Best accuracy, slowest |
Trade-offs: disk space and (for big Whisper variants) RAM. Apple Silicon handles Large v3 comfortably with 16 GB+ RAM; older machines should pick smaller.
Privacy: audio never leaves your Mac.
Cloud models
| Provider | Model | Notes |
|---|---|---|
| Groq | Whisper Large V3 Turbo | Very fast; generous free tier |
| OpenAI | Whisper-1 | Reliable, paid per use |
| Deepgram | Nova-3, Nova-2 | Lowest latency on long files |
| Gemini variants | Multilingual, paid per use |
Trade-offs: audio sent to third party, requires internet, has usage costs. Some providers (Groq) have free tiers that cover normal personal use.
Each needs an API key — see API Key Setup.
Switching models
To a local model
- Settings → Models → On-Device.
- If not downloaded, click Download.
- Click the model card to activate.
Settings → Models → On-Device tab listing Parakeet V2, Parakeet V3, and Whisper variants with download buttons
To a cloud model
- Settings → Models → Cloud API.
- Pick a recommended card or expand a provider under All models.
- If the card says Setup Required, add the API key when prompted.
If you have multiple keys for a provider, the Cloud Transcription picker has its own chooser — so you can transcribe with one key and enhance with another.
From the menu bar
The menu bar icon has a quick model submenu. Click → hover Model → pick from the list.
Language
Once a model is active, a Language dropdown appears on its card if it supports multiple languages.
Auto-detect works but adds processing time and occasionally misidentifies short utterances. Set it explicitly if you primarily speak one language.
| Family | Languages |
|---|---|
| Parakeet V2 | English |
| Parakeet V3 | 25 (incl. Spanish, French, German, Mandarin, Japanese, Hindi, Arabic) |
| Whisper (all sizes) | 99+ |
Which one to actually pick
- English speakers: Parakeet V2. Fast, accurate, private, small.
- Multilingual: Parakeet V3 if your language is in the 25, otherwise Whisper Large v3.
- Privacy-critical: any local model.
- Hour-long files in bulk: a cloud provider (Groq is usually fastest).
- No disk space: any cloud provider.
You can always change later. Downloaded models can be removed from the same panel.