Transcription & models
Choose the accuracy/speed tradeoff that fits your machine — all running locally.
Engines
- Whisper (always available) — OpenAI's model, 99+ languages, English-only
.envariants that run ~20–30% faster. - NVIDIA Parakeet (optional) — a single fast model covering 25 languages.
Whisper model sizes
Models download on first use and are cached locally:
| Model | Size | Best for |
|---|---|---|
large-v3-turbo-q5_0 (default) | ~830 MB | Balanced accuracy and speed |
large-v3-turbo | ~1.4 GB | Higher accuracy |
large-v3 | ~2.9 GB | Highest Whisper accuracy |
small.en / medium.en | ~40–140 MB | Fast, English-only, light machines |
Re-transcribe anytime
If you kept the audio, you can re-run a session with a better model — or switch between Whisper and Parakeet — without recording again. Nothing is re-uploaded because nothing was uploaded to begin with.
Live vs. after the call
Live transcription shows text as people speak. Turning it off (Settings) makes Zerds capture-only during the call and transcribe when you stop — lighter on the CPU for modest hardware.
GPU acceleration
- NVIDIA GPU (CUDA): recommended but not required. When a CUDA-capable card is present, transcription uses it; this is what makes live transcription comfortable on the larger models.
- CPU only: everything still works. Pick a smaller model, or turn off live transcription so the whole recording is transcribed after you stop — the transcript is the same, it just arrives after the call rather than during it.