WhisperDictation v1.1.0

A reliability, privacy, and performance overhaul — every subsystem reviewed, fixed, and now covered by a 112-test suite that runs in CI.

Reliability

• The microphone picker in Settings now actually selects your mic (previously it silently used the system default)
• Transcription failures, download failures, and mic disconnects surface as menu bar alerts instead of failing silently
• Model downloads are verified against pinned SHA256 hashes, with per-model progress and cancel
• Recordings cap at 5 minutes (memory-bounded) and still transcribe what was captured
• Cancel a stuck transcription with the hotkey (tap in push-to-talk, hold in toggle mode)

Privacy

• Dictated text is never logged in release builds — verify yourself: strings the binary and find no transcript output
• Docs and audit commands on the website corrected to match the source layout

Performance

• ~3.4× faster text correction (one compiled regex instead of ~230 recompiled per dictation)
• Typing no longer delays decoding; inference no longer blocks the app's thread pool

New

• First-launch onboarding: permissions walkthrough + recommended 181 MB Q5 model (existing users never see it)
• Emoji and other astral characters are no longer mangled mid-typing (surrogate-pair chunking fix)
• whisper.cpp upgraded to v1.9.1
• Supply-chain hardening: GitHub Actions SHA-pinned, least-privilege CI token

Install

Download the DMG, drag to /Applications, right-click → Open on first launch (ad-hoc signed).

SHA256 WhisperDictation.dmg:

76de1835922e2f6f4031483b0fa52a8e5c7e576dd0c3edf4a371dc31f39aaeaa

v1.0.9

v1.0.9 — Toggle hotkey mode (carpal-tunnel friendly)

A second hotkey mode alongside push-to-talk: press once to start, press once to stop, gated by a configurable hold-to-activate duration so it won't trigger by accident or in shortcut combos.

Built as a universal binary — works on both Apple Silicon and Intel Macs.

What's new

• Toggle mode — press to start, press to stop. No need to hold the key while you talk. Built for long-form dictation and anyone managing carpal tunnel / RSI.
• Configurable hold-to-activate — set anywhere from 0.5s to 3.0s (default 1.5s). Long enough to prevent accidental activation, short enough to feel responsive.
• Symmetric hold-to-stop — same hold gates both starting and stopping. The trailing silent hold is automatically trimmed from the audio so Whisper doesn't hallucinate punctuation.
• Pulsing menu bar indicator — orange circle.dotted confirms the hold is being registered.
• Settings UI — single merged "Hotkey" card with mode picker and conditional duration slider.

Default mode remains push-to-talk, so existing users see no behavior change unless they opt in.

Reliability fixes

• HotkeyMonitor watchdog now synthesizes the missed key-up when macOS silently disables the event tap mid-press, so DictationEngine recovers cleanly instead of getting stuck in .recording.
• Restored AVAudioConverter.reset() before each conversion (the converter expects a continuous stream across discrete tap-callback buffers).

Install

Download WhisperDictation.dmg below. Universal binary — runs natively on both Apple Silicon and Intel.

Verify download integrity

shasum -a 256 WhisperDictation.dmg

Expected SHA256: see WhisperDictation.dmg.sha256 attached to this release.

Full Changelog: v1.0.8...v1.0.9

WhisperDictation v1.0.8

Custom Names & Terms Vocabulary

Add your own names, places, and frequently used terms for better recognition.

What's New

• Names & Terms section in Settings → Vocabulary — add personal terms with correct casing
• Terms bias Whisper recognition AND get casing corrected post-transcription
• Flow-wrapped pill UI with add/delete
• Thread-safe cached single combined regex (one string scan regardless of term count)
• 100-term limit with word budget enforcement to stay under Whisper's token limit
• 56 tests passing (5 new)

Install

1. Download WhisperDictation.dmg
2. Open DMG, drag to Applications
3. If macOS blocks it: System Settings > Privacy & Security > Open Anyway, or run xattr -cr /Applications/WhisperDictation.app
4. Open WhisperDictation
5. Grant Microphone + Accessibility permissions
6. Settings > Model > download Small Q5 (181 MB, recommended)

Requires macOS 14 (Sonoma) or later.

Build from Source

git clone --recurse-submodules https://github.com/sam-pop/WhisperDictation.git
cd WhisperDictation
make whisper && make model && make run

Full Changelog: v1.0.7...v1.0.8

WhisperDictation v1.0.7 — Polished Release

WhisperDictation v1.0.7

Polished, stable release with all tests passing and expanded developer vocabulary.

Quality

• 51 tests passing — settings, state machine, models, TextCorrector (numbers, acronyms, capitalization, punctuation), audio devices, WhisperBridge
• All XcodeGen/xcodebuild issues resolved
• Comprehensive CLAUDE.md documenting all gotchas and patterns

Streaming Output

Text appears at your cursor as each segment is decoded — grammar correction applied per-segment in real time.

Speed

• GPU pre-warming (no first-use lag)
• Adaptive decoding (greedy for short clips, beam search for longer)
• VAD trims silence before inference

Accuracy

• 500-word developer vocabulary covering:
  • Frontend: React hooks, SSR/SSG, accessibility, design systems, CSS, animations, PWA
  • Backend: middleware, caching, auth, message brokers, circuit breakers, queues
  • DevOps: CI/CD, canary deploys, SLAs, feature flags, monitoring
  • UI: tooltip, dropdown, modal, accordion, carousel, pagination
  • AI/ML: LLM, RAG, embeddings, PyTorch, transformers
  • Languages & frameworks: 30+ languages, 50+ frameworks
• Hallucination suppression, independent transcription context
• 100+ term auto-casing (api->API, javascript->JavaScript)
• Number word conversion (three hundred->300, eighty eighty->8,080)

Install

1. Download WhisperDictation.dmg
2. Open DMG, drag to Applications
3. If macOS blocks it: xattr -cr /Applications/WhisperDictation.app
4. Open WhisperDictation
5. Grant Microphone + Accessibility permissions
6. Settings > Model > download Small Q5 (181 MB, recommended)

Requires macOS 14 (Sonoma) or later.

Build from Source

git clone --recurse-submodules https://github.com/sam-pop/WhisperDictation.git
cd WhisperDictation
make whisper && make model && make run

Full Changelog: v1.0.6...v1.0.7

WhisperDictation v1.0.6 — Streaming, Speed & Accuracy

WhisperDictation v1.0.6

Major update focused on speed, accuracy, and developer experience.

Streaming Output

Text now appears at your cursor as Whisper decodes each segment — no more waiting for the full transcription to finish. Grammar correction is applied per-segment in real time.

Speed Improvements

• GPU pre-warming — dummy inference at model load eliminates first-use lag
• Adaptive decoding — greedy for short clips (<2s), beam search for longer recordings
• VAD (Voice Activity Detection) — Silero model trims silence before inference

Accuracy Improvements

• Fixed hallucination carry-over — each transcription is now independent (no_context=true)
• Fixed vocabulary prompt overflow — previous 1200-token prompt exceeded Whisper's 1024 limit, degrading accuracy
• 500-word developer vocabulary — covers frontend (React hooks, SSR, accessibility, design systems), backend (middleware, caching, message brokers, circuit breakers), DevOps (canary deploys, SLAs, feature flags), databases, APIs, AI/ML, algorithms, and more

Bug Fixes

• Fixed audio capture producing 0 samples (converter state + engine lifecycle)
• Grammar/number correction now applied in streaming callback
• Per-engine mic selection (no longer changes system-wide default)
• TextInjector on dedicated serial queue (was starving cooperative thread pool)
• NSSound dispatched to main thread (was crashing on background)
• State machine guaranteed to return to idle on all code paths
• Comprehensive test suite (50+ tests)

Developer Tools

• /release skill — build DMG, tag, and publish in one command
• /review skill — launch parallel code review agents
• regression-tester agent — verifies changes against CLAUDE.md gotchas

Install

1. Download WhisperDictation.dmg below
2. Open DMG, drag WhisperDictation to Applications
3. If macOS blocks it, run: xattr -cr /Applications/WhisperDictation.app
4. Open WhisperDictation
5. Grant Microphone and Accessibility permissions
6. Settings > Model > download Small Q5 (recommended, 181 MB)

Requires macOS 14 (Sonoma) or later.

Full Changelog: v1.0.5...v1.0.6

WhisperDictation v1.0.5

WhisperDictation v1.0.5

New

• Number word to digit conversion: "three hundred" → 300, "eighty eighty" → 8,080
• Microphone selector in Settings
• macOS 14+ support (Sonoma and later)
• No restart needed after granting permissions — hotkey monitor auto-restarts when Accessibility is granted

Fixed

• Number conversion handles zero, digit sequences, and compounds correctly
• No trailing periods after numbers
• Deployment target matches macOS 14.0 for broad compatibility

Install

1. Download WhisperDictation.dmg
2. Drag to Applications
3. First launch: System Settings > Privacy & Security > Open Anyway
4. Grant Microphone + Accessibility permissions — hotkey works immediately, no restart needed
5. Download a model in Settings > Model (Small Q5 recommended)

Requires macOS 14 (Sonoma) or later.

Full Changelog: v1.0.4...v1.0.5

WhisperDictation v1.0.4

WhisperDictation v1.0.4

New

• Microphone selector in Settings > General — pick which input device to use for recording
• Lists all available audio input devices via CoreAudio

Fixed

• Menu bar dropdown no longer wraps text ("WhisperDictation" and status text stay on one line)
• Model badge shows friendly name ("Small Q5") instead of raw filename
• Hotkey label shortened to fit ("R⌥" instead of "Right ⌥")
• Wider dropdown (320px) for better readability

Install

1. Download WhisperDictation.dmg
2. Drag to Applications
3. First launch: System Settings > Privacy & Security > Open Anyway
4. Grant Microphone + Accessibility permissions
5. Download a model in Settings > Model

Full Changelog: v1.0.3...v1.0.4

WhisperDictation v1.0.3

WhisperDictation v1.0.3

Bug fix release addressing issues found in code review.

Bug Fixes

• Fixed race condition in AudioCapture between audio tap thread and stopRecording()
• Fixed dead code in TextCorrector that allocated and discarded regex results every transcription
• Fixed memory safety in WhisperBridge — strdup pointers tracked separately to avoid freeing static C strings
• Fixed thread safety in HotkeyMonitor — isKeyHeld now protected with os_unfair_lock
• Fixed NSEvent monitor leak — onDisappear moved to correct view in hotkey recorder
• Fixed concurrent download guard in ModelManager
• Fixed launch-at-login — now reconciles UserDefaults with actual SMAppService status on startup
• Fixed fresh clone build — make app now auto-triggers whisper.cpp build if needed
• Fixed CI cache key — was hashing non-existent submodule path, now uses .gitmodules + CMakeLists.txt
• Fixed build-whisper.sh — checks for initialized submodule before running cmake

Install

1. Download WhisperDictation.dmg below
2. Drag to Applications
3. First launch: System Settings > Privacy & Security > Open Anyway
4. Grant Microphone + Accessibility permissions
5. Download a model in Settings > Model (Small Q5 recommended)

Full Changelog: v1.0.2...v1.0.3

WhisperDictation v1.0.2

WhisperDictation v1.0.2

Speed Improvements

• Quantized models (Q5): 2-3x smaller and faster with near-identical accuracy
  • Small Q5: 181 MB (vs 466 MB full) -- recommended default
  • Base Q5: 57 MB -- fastest option
  • Medium Q5: 515 MB -- best accuracy
• Voice Activity Detection (VAD): Silero VAD model auto-trims silence before inference
• Adaptive decoding: Beam search on Apple Silicon (GPU), greedy on Intel (CPU)

Accuracy Improvements

• Temperature fallback: Retries with higher temperature on low-confidence results
• Hallucination suppression: Regex filter blocks common Whisper artifacts ("Thank you", etc.)
• Context carry-over: Uses previous transcription context for better multi-sentence accuracy
• Grammar correction: 100+ developer term auto-casing (api->API, javascript->JavaScript)

Install

1. Download WhisperDictation.dmg below
2. Drag to Applications
3. Right-click > Open on first launch
4. Grant Microphone + Accessibility permissions
5. Download a model in Settings > Model (Small Q5 recommended)

Full Changelog: v1.0.1...v1.0.2

WhisperDictation v1.0.1

WhisperDictation v1.0.1

Local, private voice-to-text for macOS. No cloud, no API keys, no subscriptions.

Install

1. Download WhisperDictation.dmg below
2. Open the DMG and drag WhisperDictation to Applications
3. Right-click > Open on first launch (required for unsigned apps)
4. Grant Microphone and Accessibility permissions when prompted
5. The app will download a Whisper model on first launch (~466 MB)

Features

• Push-to-talk dictation (hold Right Option, speak, release)
• Works in any app (Terminal, VS Code, Claude Code, browsers)
• Powered by whisper.cpp with Metal GPU acceleration
• Developer-optimized vocabulary (400+ technical terms)
• Smart grammar correction (capitalization, acronyms, punctuation)
• Multiple Whisper models (Base/Small/Medium)
• Configurable hotkey, sound feedback, dark & light mode

Requirements

• macOS 14 (Sonoma) or later
• ~500 MB disk space (app + model)

Full Changelog: v1.0.0...v1.0.1