Plainsay: native Mac dictation with WhisperKit and a reproducible benchmark

[conrader]

Hi WhisperKit community,

I’ve released Plainsay, a native macOS menu-bar dictation app. Hold Right Command, speak, release, and the transcript is inserted into the current app or kept safely on the clipboard.

Free Local mode supports Whisper through WhisperKit alongside FluidAudio’s Parakeet. The Mac client is MIT-licensed, the signed/notarized download is 12 MB, and the selected model downloads separately once.

I also published the benchmark harness, methodology, raw hypotheses, and timings. On the disclosed 50-clip LibriSpeech test-clean sample, Whisper large-v3-turbo averaged 0.68 s at 2.89% WER. The first inference is included and no hidden transcription warm-up is subtracted. This is clean English audiobook speech, not a claim about every accent, microphone, language, or room.

I’d especially value feedback on the measurement methodology, realistic corpora to add next, and any WhisperKit integration patterns that would make the local path more robust.