Transcription and Whisper models
Transcription runs on your Mac using Whisper or Apple Speech, with a choice of models and support for 15 or more languages.
Perchnote transcribes locally. You have two engines to choose from under Settings, then Audio:
- Whisper (default). Runs whisper.cpp in-process on Metal, no subprocess and no Homebrew. A Silero voice-activity gate trims silence before transcription, hallucination filtering removes the common Whisper artifacts (the "thanks for watching" tail and friends), and beam search keeps accuracy up. Works on every supported macOS version.
- Apple Speech (macOS 26 or later). Uses Apple's built-in on-device speech recognition. No model download needed, since the OS manages the language assets.
Whisper models. When you pick Whisper, you choose a model and Perchnote downloads it for you:
| Model | Label | Size | Notes |
|---|---|---|---|
base.en | Base (English) | ~148 MB | Fastest, English only, lower accuracy |
medium.en | Medium (English) | ~1.5 GB | Good accuracy, English only |
large-v3-turbo-q8_0 | Large V3 Turbo (recommended) | ~874 MB | Quantized, near lossless, fast. Start here. |
large-v3-turbo | Large V3 Turbo (full) | ~3.1 GB | Highest accuracy, largest download |
The recommended pick (large-v3-turbo-q8_0) gives you near full-size quality at a fraction of the download. Most people should use it.
Language. By default Perchnote auto-detects the spoken language. You can pin a specific language instead if auto-detect is guessing wrong. Supported languages include English, Spanish, French, German, Italian, Portuguese, Dutch, Japanese, Korean, Chinese, Arabic, Russian, Hindi, Polish, and many more. Apple Speech handles language through the OS instead.
Models are stored under your app data folder (see Privacy and data). You can switch models any time. Already-downloaded ones are detected and reused.