Draft 0.3.1

Draft 0.3.1 is the first release with the CPU path tuned as a first-class fallback, plus opt-in lists.

What changed

• CPU fallback is faster and no longer blocks startup. On a machine without a usable GPU, Draft now asks the inference session for three quarters of the logical CPUs instead of all of them. Oversubscribing the decoder measured 26 to 40 percent slower on the reference machine. The GPU-only warm-up pass is skipped on CPU, so the app is ready to dictate in about 3 seconds instead of about 12. The model, precision, decode settings, and text are unchanged.
• Live preview is honest on CPU. A preview is a full extra encode, so the switch is not offered on the CPU backend and a saved preference is turned off with a visible notice instead of doing nothing.
• Opt-in lists. A run of three or more spoken ordinals, or a make a list / make a numbered list block, becomes one item per line. Lists add only a line marker; no word is added, dropped, or reordered.
• draftctl gained --threads N, and DRAFT_THREADS pins the count. The measured CPU table and the thread policy are in docs/PERFORMANCE.md.

Installer

• File: Draft-Setup-0.3.1.exe
• Size: 1,425,649,621 bytes (about 1.33 GiB), including the F16 Whisper large-v3-turbo model
• SHA-256: D9162170756FC3AFA19793BD0AD9F0828BB3314A4F097F4F8E8DB97050BA5DF0
• Windows 10 or 11, x64 with AVX2. Per-user install under %LOCALAPPDATA%\Programs\Draft, no administrator prompt.

The installer is still unsigned, so SmartScreen may warn on first launch. Verify the hash before running it.

Known limits

• Windows x64 is the supported release. The Apple Silicon macOS preview is ad-hoc signed and has not been run on a physical Mac.
• Real-speech WER parity with the previous generation is still being measured. No accuracy claim is made ahead of it.
• The History pane is not built. The bounded, text-only store and draftctl history exist.
• There is no in-app updater and no update ping.
• On CPU, the same F16 model takes a few seconds per clip, not milliseconds. A 4 GB machine is tight because the 1.6 GB model is held in RAM; Q8_0 remains the smaller catalog alternative and is never selected silently.

Draft 0.3.0

Draft 0.3.0 shapes a finished dictation instead of pasting one block of text.

Formatting

• A paragraph break lands at a real pause: the end of a sentence, followed by silence that clears your own speaking rhythm. Draft measures that rhythm from the decoder's segment gaps and folds it into rhythm.json with a slow moving average, so it learns how you actually pause instead of guessing.
• Three break styles: Fewer breaks, Balanced (the default), and More breaks. Paragraphs can be turned off entirely.
• It never breaks in front of a continuation word such as "and" or "because", and a long monotone stretch still gets structure after 110 words.
• Formatting never changes a word. Nothing is added, dropped, reordered, or respelled. A test deletes every whitespace character from input and output and requires them to match. There is no rewrite model and there will not be one.

Spoken commands

Say these as their own phrase, with a pause on each side. The same words inside a sentence are left alone.

Say	Result
new paragraph	Start a new paragraph
new line	Start a new line
scratch that	Delete the last sentence

Commands are English only, because the rule is an English phrase list, and can be turned off.

Settings

• New Formatting section: Paragraphs with break-style chips, Spoken commands with a command reference, and a Learn my rhythm row that shows the real measured pause or says plainly that there is no measurement yet.
• The section rail is now eight sections; everything fits the fixed 680x500 window without a hidden scroll area.
• Privacy copy now counts correctly: "1 dictation stored" instead of "1 dictations".

Also

• draftctl transcribe --timings and --format expose the segment timings and the formatting rules for checking against real audio.
• Lists are deliberately not in this release. They are planned as an opt-in for 0.3.1.

Install

Draft-Setup-0.3.0.exe includes the F16 Whisper large-v3-turbo model. Windows 10/11 x64 with AVX2. The installer is not code-signed, so SmartScreen may warn. Verify the SHA-256 in the release assets first.

SHA-256: C64C71DE73285DCE8E7BA5D26FFC832C35405584A0F62CD30680BFEDA8E5A368

Draft 0.2.1

Draft 0.2.1

Private, on-device dictation for Windows. Hold Right Ctrl, speak naturally,
and release - the transcript appears at your cursor.

Download

• Draft-Setup-0.2.1.exe - Windows 10/11 x64 with AVX2, per-user install
  under %LOCALAPPDATA%\Programs\Draft; no administrator prompt.
• The F16 Whisper large-v3-turbo model is bundled in this installer. No
  separate model download is needed.
• SHA-256: d0aff26862f88f38b67ff1e1be628817145827bd10fa3d3199e3dc377d7b42c9
  (also attached as Draft-Setup-0.2.1.exe.sha256).

In this release

• A rebuilt recording HUD. One compact strip shows the real input level,
  the elapsed time, and Esc to cancel. It floats above other windows and
  never takes the cursor.
• A direct uninstall. The Start menu now has a Draft folder with
  Draft and Uninstall Draft, so removal no longer requires Settings.
• Uninstall keeps your data. It removes the app, its shortcuts, and its
  license texts. Settings, history, logs, and the model stay in
  %LOCALAPPDATA%\Draft4, and the confirmation says so.
• A cleaner upgrade. The installer clears the old top-level Start-menu
  shortcut and any stale draft.prev.exe from earlier builds.

Everything from 0.2.0 is unchanged: local Vulkan inference with the model kept
warm, the fixed 680x500 window, light and dark themes, six dictation shortcuts,
a microphone check, tray actions, launch at sign-in, and bounded local history.

Install

1. Download and run Draft-Setup-0.2.1.exe.
2. Windows SmartScreen may warn because this installer is not code-signed
   yet. Choose More info -> Run anyway if you trust the download, or
   verify the SHA-256 above.
3. Launch Draft, hold Right Ctrl, speak, and release.

Upgrading from 0.2.0: run the installer over the existing install. Settings and
history stay in place.

Good to know

Draft 0.2.1 is an early Windows x64 release and needs an AVX2 processor
(Haswell, 2013, or newer); every Windows 11 machine qualifies. An Apple Silicon
macOS preview is available separately as v0.2.0-macos.1; it is ad-hoc signed,
not notarized, and has not been run on a physical Mac. Linux is not supported.
Real-speech WER parity with the previous generation is still being measured, so
no accuracy claim is made ahead of it. The History pane is not built; the
bounded store and draftctl history exist. Updates are manual.

The model is distributed with Draft and is subject to its own terms. Full
credits and third-party notices are in the repository:
CREDITS.md and
THIRD_PARTY_NOTICES.md.
The project license is MIT.

Read the README for the full
feature list, the privacy model, and the measured performance numbers.

Draft 0.2.0 for macOS (Apple Silicon preview)

Draft 0.2.0 for macOS

Private, on-device dictation for Apple Silicon Macs. Hold Right Ctrl, speak
naturally, and release. Whisper large-v3-turbo transcribes on Metal, on this Mac.

Download

• Draft-0.2.0-macos-arm64.dmg - macOS 13 or later, Apple Silicon
  (M1/M2/M3/M4). This build is not for Intel Macs.
• The F16 Whisper large-v3-turbo model is bundled in the DMG. No model
  download is needed.
• SHA-256: attached beside the DMG as
  Draft-0.2.0-macos-arm64.dmg.sha256.

First launch

This is an unsigned preview. It is ad-hoc signed only, not notarized, so macOS
will block the first launch until you allow it:

1. Open the DMG and drag Draft to Applications.
2. In Applications, right-click Draft and choose Open, then Open
   again. If macOS does not offer Open, use System Settings, Privacy & Security,
   Open Anyway.
3. If you prefer Terminal, the equivalent is:
   xattr -dr com.apple.quarantine /Applications/Draft.app
4. Grant Accessibility and Input Monitoring when Draft asks. The exact
   panes are under Settings, Dictation, Permissions. Relaunch after granting.
5. Hold Right Ctrl, speak, and release.

Because the signature is ad-hoc, replacing the app with a new build can reset
the privacy grants. Do not move the app after granting them.

Status

This is a test build, not a verified macOS release:

• CI builds, links, bundles, ad-hoc signs, and verifies the app on Apple
  hardware, and this DMG was produced by that job.
• Nobody has dictated on a physical Mac yet. The event tap, key suppression,
  Escape cancellation, paste, floating HUD, permission onboarding, launch at
  sign-in, and Metal inference are implemented but not runtime-verified.
• Developer ID signing and notarization have not been run. Both need an Apple
  Developer account and repository secrets.
• Updates are manual; the app does not check for updates.
• Windows stays on v0.2.0 for now:
  https://github.com/TirupMehta/draft-rs/releases/tag/v0.2.0

Privacy

Audio is captured only while the dictation key is held, transcribed on this Mac,
pasted at the cursor, and discarded. No account, no telemetry, no audio upload.
Settings, history, and logs live in ~/Library/Application Support/Draft4; the
bundled model stays inside the app.

Full credits and third-party notices are in the repository:
CREDITS.md and
THIRD_PARTY_NOTICES.md.
The project license is MIT.

Draft 0.2.0 - private, on-device dictation for Windows

Draft 0.2.0

Private, on-device dictation for Windows. Hold Right Ctrl, speak naturally,
and release - the transcript appears at your cursor.

Download

• Draft-Setup-0.2.0.exe - Windows 10/11 x64, per-user install under
  %LOCALAPPDATA%\Programs\Draft; no administrator prompt.
• The F16 Whisper large-v3-turbo model is bundled in this installer. No
  separate model download is needed.
• SHA-256: 99B672AC767CA1CC8FED638CF601A81E908F97D9CA0526827BA77221F0758685
  (also attached as Draft-Setup-0.2.0.exe.sha256).

In this release

• Local Whisper large-v3-turbo inference on Vulkan, with the model kept warm.
• One fixed 680×500 window: Dictate and Settings share the same surface.
• Light is the default theme; dark is a designed second theme.
• A click-through recording HUD with the real input level and elapsed time.
• Settings for Dictation, Audio, Recognition, Appearance, Privacy, Advanced,
  and Credits.
• Credits includes the author, GitHub and X links, the open projects behind
  Draft, and a button that opens the installed license texts.
• Six dictation shortcuts, a microphone level check, tray actions, and
  launch-at-sign-in.
• Bounded local history: text and duration only.
• MIT-licensed source, no telemetry, no account, and no audio upload.

Install

1. Download and run Draft-Setup-0.2.0.exe.
2. Windows SmartScreen may warn because this installer is not code-signed
   yet. Choose More info → Run anyway if you trust the download, or verify
   the SHA-256 above.
3. Launch Draft, hold Right Ctrl, speak, and release.

Good to know

Draft 0.2.0 is an early Windows x64 release. macOS and Linux are not supported
yet. Real-speech WER parity with the previous generation is still being
measured, so no accuracy claim is made ahead of it. The History pane is not
built; the bounded store and draftctl history exist. Updates are manual.

The model is distributed with Draft and is subject to its own terms. Full
credits and third-party notices are in the repository:
CREDITS.md and
THIRD_PARTY_NOTICES.md.
The project license is MIT.

Read the README for the full
feature list, the privacy model, and the measured performance numbers.