Changelog

• Long dictations no longer lose text. A dictation that happened to say "see you in the next meeting" or "hope you enjoyed…" was thrown away as a speech-model glitch; now only short clips are checked for that. Hands-free sessions transcribe their 25–40s pieces one at a time and in order, and every piece is typed even if the last ones take a while — 15 minutes of talking comes out whole.
• Paste Last Dictation (menu bar icon): if a dictation landed where there was no text field, paste it again wherever your cursor is now. Kept in memory only, never saved to disk.
• Dictation now appears all at once, like a paste, instead of being typed out letter by letter.
• Names you correct once are now used on every later dictation: one fix (typed, or spelled out loud) also catches the name's other mishearings ("Druv", "Dhruva", "Dhruvmehra"), possessives ("Pipe's" → "Pype's"), and first and last name together ("druv mera" → "Dhruv Mehra"). Possessive fixes ("Drew's" → "Dhruv's") are learned too, and so are fixes where the mishearing sounds nothing like the name ("Python" → "Pype"), just a capital letter ("pype" → "Pype"), or a spelling said after an odd mishearing ("…Mana. That's spelled M-E-H-R-A"). Ordinary edits (a typo, a plural, a reworded Hindi phrase) are not learned.
• Unusual names now work without an OpenRouter key. A mishearing that is an ordinary word ("pipe", "through", "mera") is only changed where the sentence says a name belongs ("I work at ___", "Email ___ about…", "Mr. ___"); "the water pipe burst", "Look at pipe now", "I drove Mehra to the airport" and "mera naam…" stay as they are.
• Fixing a misheard name now teaches Airboard for a full minute after you dictate — even if you keep dictating, switch apps briefly, or fix two names. Works without an OpenRouter key too (then only clear name fixes, like "bright" → "Brite", become a card).
• Vocabulary learning now works in Electron apps (Slack, Notion, Linear, Discord, Obsidian…): Airboard asks them to expose their text to Accessibility when you start dictating into them. VS Code-family editors are left alone.
• One place for your OpenRouter key: popover → OpenRouter key. The Assistant and Vocabulary screens just show whether it's set.
• Removed AI Cleanup (cloud grammar rewrite) and the Memory voice commands ("remember…", "write my…"). Airboard is going fully local; grammar comes back later, rebuilt on-device. The Memory screen is now Vocabulary: the words Airboard has learned about you. Older memory notes stay untouched in memory.md.
• Airboard now learns your names and terms. Fix a misheard word right after dictating (or let it notice names on your screen) and a small card asks to remember it — ignore it and it's kept. From then on the word is written correctly, but only where it fits ("Ask bright AI" → Brite AI; "a bright idea" stays). Uses your own OpenRouter key; each check sends one sentence and one word. Toggle in popover → Vocabulary.
• Learning now catches more of your fixes: two names fixed in one dictation, a fix followed by more typing, and names with symbols (AT&T, C++, snake_case) — while still ignoring rewrites and grammar edits.
• Spell a name out loud and Airboard gets it right: "Tell Inakshi, I N A K S H I, that…" types "Tell Inakshi that…", and a card offers to remember the spelling. Works in any app, including terminals. Saying "that's spelled B R I T E" right after a dictation teaches the same way.

• Ask the Assistant: hold hotkey + ⌘ and ask anything — time zones, currency, quick facts. Answers by toast in seconds; one-click setup in the popover (pi engine + OpenRouter key). The assistant can fetch web pages and do exact math, and can never touch your files.
• Memory examples in the app and README now use neutral placeholder names
• Fixed: long dictations sometimes came out as only the last sentence. Saying "I mean", "no wait" or "scratch that" anywhere made cleanup throw away everything before it. That self-correction rule is gone; cleanup now only removes filler sounds (um, uh, hmm).

Added

• Airboard Memory: teach spellings ("correct pipe to Pype") and facts ("remember I work at Pype") by voice in command mode; recall facts anywhere ("write my address" types it at your cursor). Facts are lightly formatted before storing (punctuation and your taught spellings applied — your words kept as spoken; falls back to raw offline), and spellings are applied in context by AI Cleanup — "the water pipe" stays "pipe". A toast confirms every teach on screen. Memory is a flat list of plain-language lines in a markdown file (memory.md) — view and edit them as-is in the Memory window or any text editor, with a switch controlling whether they're shared with the cleanup LLM
• Memory confirmation: every "remember…" fact now shows a small card to review and edit before it's saved — and correcting a misheard name in the card teaches Airboard the right spelling automatically
• Voice editing in command mode: "delete all", "delete last word", "delete last two sentences", and "scratch that" (erases exactly what Airboard just typed) — works anywhere dictation works
• Performance window shows your recent dictations with a timing breakdown (audio length, speech-to-text ms, AI cleanup ms) from a local log that never leaves your Mac
• Anonymous performance telemetry (production builds only): timing numbers and outcome flags go to TelemetryDeck so we can see how fast dictation is across installs — never any text you dictate. "Share anonymous performance stats" in the Performance window turns it off
• Natural memory phrasing: no exact keyword needed — "don't forget that…", "note this…", "keep in mind…" all work; when a command isn't recognized, the AI interprets whether you meant to save, recall, or teach a spelling (the confirmation card still guards every save)

Changed

• Nothing is stored to memory silently anymore — facts always pass through the confirmation card

Fixed

• Dictations no longer start with a stray leading space, and chained utterances still get their separating space: the old heuristic guessed from the field's last character whenever the cursor position was unreadable (terminals always are). Now Airboard reads the character before the cursor when the app reveals it, and otherwise adds a space only when chaining right after its own previous insertion (same app, within 30s)
• AI Cleanup is now light-touch by default: it keeps your words and sentence structure as spoken — no rephrasing or synonym swaps — fixing only punctuation, capitalization, clear errors, and formatting. It also no longer invents paragraph breaks. (If you saved a custom system prompt, press "Reset to default" to pick this up)
• Installer window labels ("Airboard", "Applications") are now readable: Finder always draws DMG labels in black, so the background art gives them light plates
• Installer tagline updated to "On-device speech to text · Optional AI cleanup" — accurate now that cleanup can use a remote LLM

Added

• Styled DMG installer per the design system: drop-zone layout with drag-to-Applications arrow, version stamp, and privacy line

Fixed

• The installer DMG no longer contains a stray copy of itself (1.0.7's disk image showed "Airboard-1.0.7.dmg" inside its own window)

Fixed

• Turning AI Cleanup on now opens the setup window when the API key is missing (not only when the server URL/model are empty)
• The AI Cleanup toggle no longer stays on when cleanup can't work: it snaps back off until a valid config exists, and turns itself on when you save one
• API keys are now stored per server: a Cerebras key is never sent to OpenRouter/Ollama when you switch providers (existing key migrates automatically)
• Bluetooth earphones no longer drop to muffled call-quality audio when Airboard launches: the app no longer opens the default (Bluetooth) microphone at startup
• Recording from a selected mic actually works now: the capture core was rebuilt on AVCaptureSession. The old engine's device pinning was the root cause of every mic failure to date — truncated recordings (~0.2s of audio captured regardless of how long you spoke), hangs when recording started, and the earlier deadlocks. Recording via a chosen mic also no longer touches the Bluetooth mic at all, so music stays crisp while you dictate through the MacBook mic
• Popover logo is now the Airboard brand red (was mistakenly blue)
• Dictating a request ("can you give me three points on…") is now always transcribed, never answered or refused by the cleanup LLM: dictation is sent as tagged material-to-edit, and any refusal the model produces anyway is discarded in favor of your own words
• AI Cleanup settings now show the exact system prompt sent with every request — and let you edit it (auto-saves; one click restores the default)

Changed

• AI Cleanup settings always show which server a key is saved for ("Key saved for api.cerebras.ai"), with a Remove button
• Cerebras is the default AI Cleanup provider (preset fills gpt-oss-120b)

Changed

• Popover polish per the Airboard design system: "AI Cleanup" title casing, Microphone row uses the semantic blue accent

Added

• Microphone selection with per-device memory: pick which mic Airboard uses from the menu popover; the choice is remembered for each connected headset (e.g. keep using the MacBook mic when Bluetooth earphones connect)
• New look: the entire app adopts the Airboard design system v2 "Sleek Dark" — dark-only, tokenized surfaces, blue-led actions
• New guided onboarding for first-run setup: six steps (Welcome → The Gesture → Microphone → Accessibility → Your Hotkey → Try It), skippable at any point; existing users never see it

Changed

• AI cleanup now defaults to OFF — a fresh install has no server configured, so an on-looking toggle was misleading; flipping it on still opens the setup window when unconfigured

Fixed

• The AI cleanup settings window could not be closed reliably: its transparent chrome let clicks pass through to the desktop, triggering macOS's show-desktop (all windows hiding). It now uses a standard title bar
• Opening cleanup settings no longer triggers a Keychain permission prompt when dev and prod builds coexist — each build now stores its API key under its own Keychain entry

Fixed

• Released builds could not access the microphone (no permission prompt, no entry in System Settings): the release pipeline's re-signing step was stripping the app's entitlements. Affected 1.0.2–1.0.4; the pipeline now verifies entitlements and signature instead of re-signing

⚠️ Do not install this version — microphone access is broken in this build (a release-pipeline bug stripped the app's entitlements). Install the latest release instead; if you already installed 1.0.4, it will auto-update itself.

Added

• MIT LICENSE file (the README already claimed MIT; now it's true)
• Automatic updates (Sparkle): the app checks at launch and daily, downloads in the background, and installs on quit. "Check for Updates" available in the menu popover

Changed

• AI cleanup now runs on dictations of 6+ words (was 12) — fast providers like Cerebras made the round-trip imperceptible
• Performance metrics measure transcription time only, no longer including the AI cleanup round-trip

Added

• Filler-word removal ("um", "uh", "ah") in all dictation modes — local, always on
• Optional AI cleanup via any OpenAI-compatible endpoint (OpenRouter, Cerebras, AWS Bedrock, Ollama, vLLM): grammar and punctuation fixes, paragraph breaks, spoken enumerations formatted as bullet or numbered lists. Configured in the menu popover; API key stored in the Keychain; falls back to local rules if the server is slow or unreachable. See docs/cleanup-server-recipes.md.

Changed

• Speech engine swapped from Whisper large-v3-turbo (WhisperKit) to NVIDIA Parakeet TDT 0.6B v3 (FluidAudio): better English accuracy and ~10× faster transcription. Requires Apple Silicon.
• Multi-line and special-character text inserts via paste (fixes doubled bullets in editors with auto-list continuation)

Removed

• Custom vocabulary feature (superseded by AI cleanup)

Requires macOS 14+ on Apple Silicon. First launch downloads the ~460MB speech model; everything is offline after that unless you configure an AI cleanup server.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PiTFMhGewVKAgbJ8fBW6gz