Konrad Sierzputowski develops Plainsay, an open-source dictation application for the Mac. His work connects on-device speech recognition with the small interaction details that make transcription useful inside other applications.[1][2][3]
Start with How I benchmarked on-device speech-to-text on Apple silicon →
Links
GitHub Plainsay DEV article Contributions
Make local speech recognition fit a writing workflow
Plainsay uses a push-to-talk interaction: hold a key, speak, then insert the transcription into the application in use. Its local mode runs speech engines on the Mac after model download. Optional cloud features have a different data path; hosted transcription or text polishing should not be confused with wholly local processing.[2][3]
Publish measurements that can be rerun
Sierzputowski compared Whisper and Parakeet using the same engine implementations as his app. His benchmark publishes a small LibriSpeech sample, reference transcripts, outputs and per-clip latency. He separates inference timing from model download and preparation, and inspects individual errors rather than reporting only an aggregate score.[1]
Ideas
The comparison measures two engines on specified Apple hardware, not all dictation applications or all speech conditions. Its small sample enables manual error review, but a few words change the reported error rate. Those limits matter when assessing a different accent, microphone or environment.[1]
Tools, writing & talks
Writing, research & talks
Start here
A reproducible comparison of the Whisper and Parakeet engines used in Plainsay.
Read the dataset, timing method and limitations before the score table. Compare the disclosed setup with your own dictation needs; the committed audio and outputs make the claims inspectable without relying on a product leaderboard.
Author[1]An open-source Mac dictation app with on-device speech engines and optional cloud features.
Creator and maintainer[2]Original source