Kekoso is a macOS menu bar app that turns speech into text without sending audio anywhere. Press a hotkey, talk, and the text lands in whatever field you were typing in.

What makes it different is the engine layer. Three families of speech models run locally behind one protocol, and you pick which one is active: Parakeet TDT v3 and SenseVoice through FluidAudio, and Whisper through WhisperKit on the Neural Engine. Coverage is why this matters: Parakeet is the fastest but handles 25 European languages, SenseVoice covers Chinese, Cantonese, Japanese and Korean, and Whisper carries the long tail of 99. Models are downloaded once into Application Support, so you only keep what you use.

The hardest problem was silence. Whisper was trained on subtitled video, so on a pause it writes what subtitles usually say there: sign-off lines and translator credits. Segments with no audio behind them are dropped before anything reaches your text field, in one shared place for dictation, files and recordings.

Beyond dictation it transcribes audio and video files, YouTube links and recorded calls, exports TXT, SRT and VTT with timecodes, watches folders for new files, and applies a custom dictionary for names and terms in post-processing.

The app talks to the network in three places only: licence checks, updates, and fetching audio when you paste a YouTube link. Dictation itself works with Wi-Fi off, including the first run.

Built in Swift and SwiftUI, layered behind protocols with pure decision logic kept apart from the hardware so it can be tested: 145 source files and 1444 tests. Interface localised into ten languages. Distributed outside the App Store as a notarised build with Sparkle updates. The licensing backend is Cloudflare Pages Functions over D1, and the site is Astro.

macOS 14.4 or later, Apple silicon. $29 once for three Macs, with a 7-day trial.