What’s new in Eidetic

Added

• Pick your push-to-talk key: right Option stays the default, and Settings now offers left Option, right/left Control, right Command and fn/Globe.
• Goals on the Insights dashboard — a talk-share target, a weekly hours budget, and archive milestones, all judged on the week you can still change.
• “How you talk”: typical and longest monologue, the pause before you answer, interruptions and questions per hour, hesitation sounds per thousand words — measured locally, with published coaching benchmarks named next to them.
• The dictation dictionary learns multi-word and cross-script fixes on the second occurrence — “open router” → OpenRouter, «Энэйтэн» → n8n. Learned entries are marked with ✨.

Fixed

• The instant draft after a recording now shows the same timestamped, speaker-labelled utterances as the final transcript.
• Confirmed live lines no longer start with the punctuation of a sentence the recognition window just cut.
• Two log lines that carried short snippets of meeting text now carry lengths only, and a privacy lint keeps that whole class out of the logs.

Added

• The dictation capsule lives above the Dock — a sliver at rest, two labelled buttons on hover, and a full panel with waveform and live transcript while you dictate. Clicks never steal focus from the app you are dictating into.
• Insights: hours recorded, which days and hours they land on, a six-month heatmap, recurring topics, talk share, speaking rate and silence — computed once per meeting and cached.
• Dictation clean-up passes: hesitation sounds dropped by default, opt-in spoken punctuation behind an activation word, and a dictionary for terms the recogniser keeps mishearing.
• The dictionary teaches itself from corrections you make in the twenty seconds after a dictation.
• Dictation can be told its language, and the app names what auto-detect costs in latency instead of implying it is free.

Fixed

• A long dictation is no longer cut short, and never arrives as whatever was on your clipboard — Eidetic waits for the words to land before restoring it.
• The recogniser’s unknown-token marker no longer reaches the pasted text.
• One calm offline state instead of four separate network errors.
• An interrupted model download resumes instead of starting over.

Added

• Dictation’s final text now comes from one offline re-decode of everything captured, so a sentence next to a pause is no longer dropped.
• Live preview while you are still speaking — first text about a second in, instead of silence until you let go.
• Onboarding asks which language your meetings are in, before anything is downloaded.
• Speech engines are named and their limits stated: Parakeet, Whisper, and Apple’s own recogniser on macOS 26+.

Fixed

• An engine’s language gap no longer turns into a failed meeting.
• Whisper model size is chosen against the machine’s RAM instead of a hardcoded default.

Fixed

• ⌘⇧R and the tray’s Start/Stop item toggle recording again.
• A dead webview is recovered by navigating, not reloading — no more white window.
• Short recordings get a summary instead of a false “too short” refusal.

Fixed

• The onboarding download that could freeze at 5% forever now has a stall bound, full logging and a free-disk precheck.
• Summaries of very short recordings stop echoing the transcript back at you.

Fixed

• A recording interrupted by a quit or a crash is recovered from its journal instead of being lost.
• Diagnostic reports are consented, previewed byte-for-byte before sending, and scrubbed of meeting content.