Transcribe with Whisper instead of Parakeet

Rooster's standard engine, Local Parakeet, is fast and covers English and European languages. For other languages, use Local Whisper. It also runs entirely on your Mac.

Set it up

1. Open Settings and choose Transcription.
2. Set Transcription provider to Local Whisper.
3. Download a Whisper model file. Get Whisper models, now in Settings, opens a page that has them. Rooster recommends ggml-base.en.bin for English, or ggml-small.bin for other languages.
4. Under Whisper model, press Choose… and pick the file.

Choosing the file with Choose… lets Rooster keep reading it after it restarts. If Rooster can no longer reach it, transcription stops and asks you to choose it again in Settings.

Set Language to the language of your recordings, or leave it on Auto-detect.

Faster on the Neural Engine

Settings shows whether Whisper is using your Mac's Neural Engine. Without it, Whisper runs on the graphics processor and the CPU. To use the Neural Engine, put the matching Core ML encoder - a folder ending in -encoder.mlmodelc, from the same page - beside the model file.

How it differs from Parakeet

• Whisper covers more languages. Choosing a language Parakeet does not cover, with Parakeet selected, gets a note telling you to switch to Whisper.
• Whisper reports the time of each word, so cutting, captions and the glossary work as they do with Parakeet.
• Rooster downloads Parakeet itself; Whisper's model is a file you provide.
• Parakeet writes down filler words such as "um" for you to remove. Rooster runs Whisper in a way that leaves most of them out, so the filler-word tool finds few or none and suggests switching to Parakeet. For languages other than English it finds none with either engine.

If no model is chosen, transcription stops and says to choose a Whisper model in Settings.