Skip to content

Latest commit

 

History

History
23 lines (22 loc) · 5.16 KB

IDEAS.md

File metadata and controls

23 lines (22 loc) · 5.16 KB

Ideas (running list, newest at the bottom)

  • 10/30 — the two MacWhisper features, nothing else for v1.
  • 10/30 — engine: "whichever one is the most accurate." Add WhisperKit large-v3-turbo as a switch and compare it with Parakeet v2 on Eric's own dictation; let his voice settle it, not the leaderboard.
  • 10/30 — never lose a dictation to a focus grab (MacWhisper's worst habit): hold the text when the target app lost focus, keep it until it is pasted. (In v0.1 as decision 4.)
  • 10/30 — a toggle mode for long dictation (double-tap Right Option locks recording on, as Hex does).
  • 10/30 — a short in-memory history of recent transcripts in the menu, beyond the last one. Declined 10/31 (Eric: "I don't want history. I just like having the most recent… in the memory of the app in case something steals the focus"): the last one, in memory, is the design.
  • 10/30 — when a held transcript is waiting, paste it on the next return to the original app (or a key), instead of only from the menu.
  • 10/30 — a click mid-dictation must never throw the audio away ("another thing popped up and I had to click a mouse while I was dictating"). Fixed the same evening: a cancel only in the first 0.5 s.
  • 10/30 — real-time dictation: see the words as they're spoken, not only after release (FluidAudio's streaming managers: SlidingWindowAsrManager, StreamingEouAsrManager).
  • 10/30 — an indicator at the cursor or the text location while listening, so he can see where the words will land (Accessibility: the focused element's caret bounds; fall back near the mouse for apps that don't expose it).
  • 10/30 — live vs. one blob: "I don't know if I want an option to have it live transcribe or just paste one blob at the end." v0.2 shows the live words in the bubble and pastes one blob on release, with a menu toggle for the live preview. Typing the live words straight into the field is a possible third mode, after he has lived with the bubble.
  • 10/30 — the bubble: "I like it… it's actually also showing it in the bubble. Really neat."
  • 10/30 — meeting transcription from the menu (feature 2, grown): "multi speaker live meeting transcription is easier we just do it afterwards. A bit of both would be nice." A simple UI from the menu, and the icon changes while it records a meeting and while it transcribes. Shape: live rough text during, a speaker-labeled pass after (FluidAudio's Diarizer trait, off in v0.1). Open: calls need the other side's audio (system audio capture), not just the mic; where meeting recordings are kept.
  • 10/30 — "Do we have an icon for it, and a UI, an about page, a settings page?" None yet: the menu bar uses the stock SF mic (easy to confuse with macOS's own mic indicator), no app icon, no About, no Settings. Proposed: a feather/nib mark that changes with state plus an app icon; a Settings window (hotkey, live preview, launch at login, later the engine and the meeting folder); an About panel (version, engine/model, FluidAudio and Parakeet credits).
  • 10/31 — the bubble as the app icon with the words streaming out to its right on one line, older words scrolling off; no red dot, no "Listening → app" (Eric). Built the same night.
  • 10/31 — the words as a smooth stream: soft grey, sliding across, fading out on the left; the bubble with a glowing edge (Eric). Built the same night.
  • 10/31 — go public once a few more basic features are in ("I might as well make this public… once I add some more basic features"). Before then: a name (Quill is taken by other voice apps, see CLAUDE.md), notarization (one notarytool step with the Developer ID), a license for the code, the credits generalized (About says "Local dictation for Eric"), the Parakeet CC-BY-4.0 attribution kept, a README for strangers, and a first-run window that explains the two permissions.
  • 10/31 — meeting mode v0.3: record, then offer to transcribe with Whisper Large v3 Turbo (Eric: "build in the meeting recording feature, which then offers to transcribe it with… Whisper Large Turbo"). Me/Them from the two tracks.
  • 10/31 — "maybe the recording just needs to be like a real-time transcription thing, or perhaps not, because transcription isn't real-time, I thought." It can be: Parakeet runs ~100x faster than speech (the bubble already does live text). A live view during the meeting, with Whisper Turbo after for the accurate copy, is a later layer.
  • 10/31 — the bubble as a small, subtle, dark Apple-style HUD, not white (Eric). Built. Also floated: a button to start a meeting recording from a pop-up; he chose "just the bubble" for now, so a popover interface from the feather is parked here.
  • 11/14 — publish it, free ("it feels like a better lightweight version than some of the other whisper transcription apps for OS X. I would just make it free"). Already: public, MIT, notarized release. Still needed: a name (humanitas-labs/quill, ~4k stars, is a feather-icon meeting recorder), first-run onboarding for the permissions, a landing page, a Homebrew cask, a launch post.
  • 11/14 — typing live into the field ("Type as I speak"): considered and declined (Eric: "leave it off - sounds like it's not worth it"). Revisions mean backspacing into his app, and it breaks the never-paste-into-the-wrong-app rule.