Changelog
What’s new in DictateSI.
Every shipped release, what changed, why it matters. The
source-of-truth document lives in the app repo as
CHANGELOG.md;
this page mirrors it.
Only what we are licensed to ship. The optional Qwen2.5 3B cleanup model is gone: its license permits
research use only, which our notices had missed. Everything DictateSI embeds or downloads is now stated
exactly (MIT or Apache-2.0, nothing else) and a build test keeps it that way. Also in this release: a
double-tap keeps a recording open, license keys arrive and change nothing during the free beta, and the
signed binary carries one entitlement fewer.
- • The 3B model is gone. It was never the default and never bundled. A saved pick of it falls back to the default model, and the file an earlier version downloaded is removed at the next launch, giving 2.1 GB back: the app now sweeps its own cleanup-model folder of files no current model can load, and touches nothing else.
- • Notices exact, and guarded. whisper.cpp and llama.cpp (MIT, at pinned commits), the Whisper model files (MIT), Qwen2.5-1.5B-Instruct (Apache-2.0) and DictateSI's own cleanup model, a derivative of it (Apache-2.0). Four provenance tests fail the build if a code file names copyleft or competitor-derived code, if a downloadable model lacks an exact license entry, or if a model download loses its pin.
- • Tap twice to keep recording. With a modifier-only hotkey such as Fn, a quick double-tap keeps the recording open with nothing held; one more press finishes it, then cleanup runs. Settings warns when macOS also acts on the Fn key.
- • License keys, verified on the machine. A key is a signed statement (licensee, tier, seats, dates) checked against a public key built into the app: no account, no activation server, no network, and the clock is never consulted. Teams can deliver one key to a whole fleet with a configuration profile; the sample profile is in the deployment kit. Nothing is gated: DictateSI stays free during the beta and a key changes no behavior.
- • One entitlement fewer. The unused Apple Events entitlement is gone from the signed binary, and the privacy tests now also cover the recording and transcription path. 176 tests.
Your clipboard comes back. Pasting a dictation used to leave it on the clipboard, so a later ⌘V pasted
the dictation instead of what you had copied. The app now snapshots the clipboard (every item and type:
text, rich text, images, files), pastes, and restores it 600 ms later once the target app has read it. A
copy you make inside that window wins. Read Aloud's selection grab uses the same snapshot.
Build hygiene. Shipped builds no longer carry the debug entitlement that lets a debugger attach to the
running app. The build tool adds it by default; the shipping targets now turn it off. Nothing else changed.
Silence pastes nothing. Holding the hotkey without speaking used to paste "Thank you.", Whisper's stock
answer to silence. Silent or near-silent recordings no longer reach the speech model at all, and its known
inventions ("you", bracketed markers, "Thank you." from quiet audio) are discarded instead of pasted. A real
"Thank you." said at normal volume still goes through.
If the model picker had ever been touched before 0.8.0, the stored pick was the old default and quietly
blocked the upgrade to the DictateSI model. A one-time migration clears it; a deliberate pick of the larger
model, or any pick made from now on, is kept.
The on-device model, trained for dictation. The default cleanup model is now DictateSI's own: Qwen2.5 1.5B
fine-tuned on dictation, hosted as a tagged GitHub Release and pinned by SHA-256 like every other model.
Same size and speed as before.
- • What changes in the output. Fillers, stutters, and false starts go. Money, percentages, dates, times, dosages, and ages are written the standard way ($4,217.50, 18%, October 16, 2:30 p.m., 500 mg, 62-year-old). A spoken self-correction keeps only the corrected value. Shorthand whose magnitude was not said ("up from one seventy five") stays as spoken. Every name and figure is kept exactly.
- • Measured against a written gate. 37 of 40 exact on the primary set (the stock model: 12 of 40) and 18 of 23 on a held-out set that includes three long paragraphs, with zero invented or merged figures, zero lost names, zero leftover fillers or correction cues, and zero dropped content on both. Seven training runs were rejected on the way for exactly those faults. Details under Accuracy, measured.
- • No interruption on upgrade. Cleanup keeps using the model you have while the new one downloads in the background, then switches at the next dictation. The general-purpose models remain selectable.
A stutter is not content. The 0.7.1 guard counted a repeated phrase ("the Third Circuit, um, the Third Circuit")
as content and so rejected a correct cleanup; repeats now collapse before counting. Found by a second,
held-out evaluation set.
Never lose a sentence. The first measured pass over the on-device cleanup model caught it returning a
paragraph minus its first sentence. Cleanup now has a content guard on every backend.
- • Content guard. A model's output is compared with what you said, counting content rather than words: fillers and stutters do not count, and a spoken number and its written form ("four thousand two hundred seventeen dollars", "$4,217") are one unit each. An output that keeps too little is discarded and your raw dictation is kept instead. Applies to the on-device model, Ollama, Anthropic, and OpenAI-compatible endpoints.
- • Measured, in the open. The app repo now carries a 40-case dictation evaluation set with documented conventions and a harness that scores any model through the real pipeline. The shipped 1.5B model scores a word error rate of 0.138 against a careful editor; the 3B model is marginally better on average but changed numbers in two cases, so it stays optional. Method and numbers: Accuracy, measured.
Better ears. OpenAI's Whisper large-v3-turbo model joins the picker and becomes the default on
Apple Silicon machines with 16 GB or more: the biggest single accuracy step available to an on-device app.
- • Hardware-tiered default. 16 GB+ Apple Silicon gets large-v3-turbo (547 MB, 5-bit quantized). Intel machines stay on small.en, 8 GB machines on base.en. Settings marks the recommended model for your machine; your choice always wins.
- • Upgrade without interruption. The app keeps dictating with the model already on disk while the new one downloads in the background, verifies its SHA-256, and takes over at the next idle moment.
- • Measured, not promised. About half a second per utterance on an M5 Pro, against 0.38 s for small.en. Better sentence punctuation; a spoken "question mark" becomes "?". Turbo writes spoken lists as "Number 1. Banana. 2. Kiwi." and the formatter now reads that form too.
- • Same rule as every model download: pinned to an immutable revision and SHA-256-verified before install. Turbo is multilingual but DictateSI runs it in English.
Intent refinements from the first real dictations: "the following items: one apple, two bananas,
three mangoes" is now a list even without saying "number one," and spoken punctuation works.
- • A list introduction relaxes the markers. After "the following," "as follows," "here are," or a colon, "one…, two…, three…" are items, not quantities. Without an introduction, "I bought one apple and two bananas" stays prose.
- • Spoken punctuation. Say "colon," "comma," "period," "question mark" and you get the mark, not the word. Real words in a sentence ("the trial period") are untouched.
- • Conjunctions spoken inside lists are dropped from items, and the on-device model is told to keep spoken list markers as spoken.
The formatting release. Say the structure and you get the structure: "number one, two, three"
pastes as a numbered list under its introducing line, "bullet point" makes bullets, and "new paragraph"
does what it says. On every backend, including with cleanup off.
- • Structure comes from code, not from the model. A deterministic formatter lays out spoken lists, bullets, and paragraph breaks and never changes your words. It is conservative on purpose: "I have two cats" and "the first quarter" stay prose.
- • The small on-device model polishes wording only, block by block, and is never asked about layout. We measured it: given layout instructions it dropped words; given one item at a time it is accurate and fast.
- • Cloud models get the full brief. Anthropic and OpenAI-compatible backends receive formatting instructions, so connecting a Sonnet-class model gives Sonnet-class documents.
- • Type-out mode sends real Return keys for line breaks. New General setting to turn formatting off.
The IT release. On-device stays the default; this gives an institution's IT team real control
over it, and gives anyone the option to plug in the AI account they already have.
- • Managed preferences: configuration profiles from Jamf, Kandji or Intune can pin cleanup to on-device, allow-list backends, force the audit log on, turn local history off, or lock an approved endpoint. Locked controls read "Managed by your organization." Policy can only narrow toward on-device. Sample profiles ship in the deployment kit.
- • Connect the AI account you already have: a new OpenAI-compatible backend for OpenAI, Azure OpenAI, Google Gemini, Groq, Mistral, OpenRouter, or a gateway your IT team runs. Pick a provider, paste the endpoint and model, add a key if needed, test the connection. Anthropic keeps its own tab.
- • Honest audit records for it: endpoint runs are logged as
endpoint, and the off-device flag comes from the host, so a gateway on this machine counts as on-device and anything else counts as off-device, even when the request fails.
The rename release. The product formerly called Ghost is now DictateSI, at dictatesi.com.
Nothing about how the app works changed; its identity did, along with a careful upgrade path.
- • New bundle identifier
com.dictatesi.mac. macOS treats this as a new app, so on first launch it asks again for Microphone and Accessibility. Grant both once. - • Automatic migration on first launch: settings, downloaded models, and the History and Audit stores move to the new identity. No re-download, and the audit hash chain is untouched.
- • Not migrated, on purpose: an Anthropic API key stored in the Keychain by the old build. Paste it again under Settings → AI Cleanup if you use the cloud backend.
- • App icon: the same ink tile and glyph as this site.
The efficiency release. DictateSI is now a polite guest on 8 GB machines, with
hardware-aware defaults, GPU-accelerated cleanup, and a strict CPU budget.
- • Hardware-tiered defaults: 8 GB-class machines default to the lighter Whisper base.en model and free the cleanup model (~1.4 GB) after 5 idle minutes; it reloads in about a second when needed. 16 GB+ machines keep the fastest setup. A new "Keep model in memory" toggle makes the trade yours.
- • GPU-accelerated cleanup on Apple Silicon (Metal, unified memory, no extra RAM). Cleanup runs ~5–7× faster and no longer spikes the CPU cores your real work is using.
- • Burst-thread budget: transcription and cleanup are capped near the performance-core count instead of grabbing every core. No fan spin-ups, no stalling the app you're dictating into.
- • The optional 3B model is marked "not recommended" on 8 GB machines.
The trust release: DictateSI's privacy claims stop being promises and become
evidence you can export, verify, and hand to a security reviewer.
Tamper-evident audit log
- • Settings → Audit Log (off by default): every dictation appends a hash-chained record of which cleanup backend ran and whether anything left the machine. Metadata only, never the transcript.
- • Editing, reordering, or deleting a record breaks the chain; the built-in Verify integrity check detects and reports it.
- • Export the whole log to JSON or CSV with an integrity summary. The headline figure for a clean deployment: off-device events: 0.
- • The log reflects what cleanup actually did, not what was configured: a failed cloud request still counts as off-device; a local no-op never does.
Verified model supply chain
- • Every AI model download (Whisper + cleanup) is now pinned to an immutable revision and SHA-256-verified against a published checksum before install. A file that fails verification is deleted, never loaded.
- • The test suite (31 tests at the time, including the privacy invariants) runs in CI on every push; no build ships without passing the moat guards.
First-run experience
- • Onboarding now offers the on-device cleanup model up front (one-time ~1.1 GB download with progress) so it's never a surprise. Enabling is always an explicit tap, and an existing cloud configuration is never overridden.
The pipeline is now fully local: dictation is transcribed and polished
entirely on your machine, with zero network and zero keys. No competitor offers frictionless
end-to-end on-device AI dictation.
On-device AI cleanup
- • Settings → AI Cleanup → On-device downloads a small open-licensed model once (Qwen2.5-1.5B, ~1.1 GB, Apache-2.0) and runs it in-process via embedded llama.cpp. No Ollama, no API key, works offline. ~1–2 s on Apple Silicon.
- • On-device is the new default for fresh installs; existing key-holders keep Cloud. "Nothing leaves the machine" is now the default cleanup posture.
- • Optional higher-quality Qwen2.5-3B download for 16 GB+ machines; in-Settings on-device test.
- • Automated moat guard: tests fail the build if any networking code enters the on-device cleanup path.
DictateSI is now bidirectional voice: dictate text in, and listen to text read
back, both entirely on-device. A proofreading and accessibility capability the cloud-only
competition structurally can’t match.
Read Aloud: on-device text-to-speech (new)
- • Select text in any app and press ⌃⌥R to hear it read back; press again or Esc to stop. Rebindable in Settings → Read Aloud.
- • On-device voices via Apple’s speech synthesizer (default, Enhanced, Premium, Personal Voice): voice picker with live preview, plus adjustable speed and pitch.
- • A calm teal speaking overlay, distinct from the recording pill.
- • Auto-stops the read when you start a dictation, so you never dictate over the voice.
- • Nothing leaves your machine: the selected text is never transmitted, never written to disk, never saved to history. No new permissions, no new network endpoints.
History controls (compliance hardening)
- • "Save dictation history" toggle: off means nothing new is recorded.
- • Retention window (forever / 1 / 7 / 30 / 90 days) with automatic purge. History never leaves the machine either way; this bounds what exists at rest.
Under the hood
- • New automated test suite, including a privacy-invariant scan that fails the build if networking or persistence code ever enters the read-aloud path.
- • Global-hotkey handler now filters by hotkey ID, so a second registered hotkey no longer re-triggers the first.
- • Type-out output mode no longer blocks the UI on long dictations.
- • Onboarding reflects your actual hotkeys and introduces Read Aloud.
The v0.2 release turns DictateSI from a working demo loop into a real
product surface: full Settings, AI cleanup, vocabularies, per-app
rules, history, and a distribution pipeline.
Distribution & onboarding
- • First-run onboarding wizard (Welcome → Microphone → Accessibility → Try it → Done).
- • Full Settings window with sidebar nav: About, Permissions, General, Audio & Model, AI Cleanup, Vocabulary, Per-App Rules, History.
- • Menu-bar quick settings: mode indicator (🔒 Local / ☁ Cloud cleanup) plus inline toggles for “Polish with AI” and “Show recording overlay.”
- • Notarized-DMG distribution pipeline producing
DictateSI-X.Y.Z.dmg. - • Privacy & data flow doc covering exactly what stays local and what goes to the cloud.
Recording
- • Custom hotkey recorder: bind any key combo or hold any modifier alone (Fn, Right ⌘, etc.).
- • Double-tap-to-lock for hold-mode hotkeys: quick double-tap latches recording on for hands-free long-form dictation; tap once more to stop.
- • Notch-style overlay with live level bars, elapsed timer, and lock indicator when latched.
- • Output mode: paste (⌘V) or type-out per-character keystrokes for password fields and apps that block paste.
- • Optional start/stop sound effects.
- • Auto-stop on silence (configurable 1–5 s). Suppressed during locked recording.
- • Esc cancels mid-recording without pasting.
Transcription
- • Whisper model picker (tiny.en, base.en, small.en) with in-app download and live model swap (no relaunch).
- • Domain vocabulary packs: Banking, Insurance, Legal, Real Estate, Title. ~20–25 terms each.
- • User-editable custom vocabulary, fed to Whisper as
initial_prompt for recognition biasing.
AI cleanup (optional, opt-in)
- • BYOK Anthropic Messages API integration; API key in macOS Keychain only.
- • Model picker with curated presets (Opus 4.7, Sonnet 4.7, Sonnet 4.6, Haiku 4.5) plus a Custom model ID field for any current Anthropic model string.
- • Three tone presets (neutral, formal, casual).
- • Custom system prompt override for vertical-specific tuning.
- • Hosted-cleanup tier scaffolded in the UI (server arrives in v0.3).
- • Graceful fallback to raw Whisper on any cleanup error; it never blocks the paste.
Per-app rules
- • Captures the frontmost app at recording start so rules bind to the target app even if focus shifts.
- • Per-app overrides for AI cleanup on/off, tone, and vocabulary pack set.
History
- • Local SwiftData persistence of every successful transcription (timestamp, duration, source app, raw text, polished text, cleanup-used flag).
- • Searchable list with copy-to-clipboard, delete, clear-all. No syncing; database lives only on this machine.
Fixed
- • Critical: audio capture was silently dropping ~99% of recorded audio because the AVAudioConverter input block signaled
.endOfStream after every tap callback. Switched to .noDataNow and added converter.reset() at recording start. - • Three residual “Quill” strings (pre-rebrand name) cleaned up in AppDelegate and entitlements.
0.1.0
2026-05-26 · pre-rebrand
Project scaffold under the working name “Quill”; rebranded
to DictateSI the same day. Never publicly distributed; listed for completeness.
- • macOS native Swift menu-bar app, XcodeGen + Makefile + stable Apple Developer signing.
- • ⌃⌥Space global hotkey via Carbon
RegisterEventHotKey. - • AVAudioEngine recording pipeline (16 kHz mono Float32 PCM).
- • whisper.cpp xcframework integration with ggml-small.en.
- • End-to-end loop: hotkey → record → whisper → paste at cursor.