SAID 0.4.0 macOS preview

SAID 0.4.0 for macOS

This is the first macOS build of SAID. It packages the existing local
SenseVoice, punctuation, VAD, and deterministic Clean pipeline in a native
menu-bar app for macOS 12 or later.

Before opening

This preview is ad-hoc signed because the project does not yet have an Apple
Developer ID certificate. Verify the DMG against SHA256SUMS-macos.txt, drag
SAID to Applications, then Control-click SAID and choose Open. If macOS
still blocks it, choose Open Anyway under System Settings → Privacy &
Security. The release is deliberately marked as a GitHub prerelease until a
notarized build is available.

Included

• Universal Apple Silicon and Intel application in a drag-to-Applications DMG.
• Control–Option–Space global start/stop shortcut.
• Local, one-shot Chinese/English dictation with optional Japanese/Korean.
• Exact and Clean output modes.
• Paste-at-caret through macOS Accessibility, with clipboard fallback.
• Verified model bundle, bundled licenses, ad-hoc local signing, and optional
  Developer ID signing and notarization.

Not yet included on macOS

• Live phrase typing.
• Adapt and its optional Qwen model.
• Per-application profiles.
• The visual setup window, login startup setting, and shortcut recorder.

Microphone audio and recognized text stay on the Mac. The application makes no
cloud transcription request.

SAID 0.3.0

SAID 0.3.0

SAID 0.3.0 turns the local voice keyboard into a controlled dictation workflow:
choose Exact, Clean, or app-aware Adapt output, optionally type stable phrases
while speaking, and keep every rewrite local to the PC.

Highlights

• Exact preserves recognizer text, Clean applies bundled deterministic
  speech repair, and Adapt uses an optional local Qwen3 0.6B model to shape
  the result for chat, mail, documents, developer prompts, editors, or shells.
• Type while I speak can insert settled phrases after natural pauses. Clean
  and Adapt revisions replace only SAID-owned text; focus or caret uncertainty
  safely moves the final result to the clipboard.
• Adapt is lazy-loaded, cancellable, removable, protected by content-preservation
  checks, and falls back to the complete Clean draft when a rewrite is unsafe.
• Setup & settings has a permanent four-section navigation rail with immediate
  saving, clearer state labels, accessible control names, and polished light,
  dark, High Contrast, reduced-motion, and 100–200% DPI layouts.
• Shortcut settings offer Right Alt, F8, and a safe custom recorder with
  rehearsal and validation for reserved or unsafe combinations.
• Recognition languages now appear as a whitelist in Dictation settings.
  Chinese and English stay enabled by default; Japanese and Korean can be
  switched on independently and persist across restarts.
• Disabled Japanese/Korean SenseVoice output is blocked before punctuation.
  Enabled Japanese/Korean text bypasses Chinese/English cleanup and Adapt so
  its original writing is preserved.
• The tray can override automatic application classification, including a
  conservative Shell fallback for terminal contexts.

Quality and safety

• Native contract, shortcut, language, refinement, performance, microphone,
  transcription, Unicode replacement, preview-isolation, and real-model suites
  passed on Windows.
• The real-model gate covers Clean and application-aware Adapt output, bilingual
  content preservation, prompt injection, long input, fallback, and a warm
  inference budget under three seconds.
• Portable and installer contents are verified to contain the four-file speech
  bundle and exclude the optional Adapt model and research caches.

Distribution

• SAID-Setup-0.3.0.exe — per-user Windows installer.
• SAID-windows-x64-0.3.0.zip — portable Windows x64 package.
• SHA256SUMS.txt — SHA-256 checksums for both artifacts.

The installer and executable are not Authenticode-signed. Windows may show a
SmartScreen warning; verify the SHA-256 checksum before running the package.

SAID 0.2.0

SAID is a lightweight Windows voice keyboard with fully local English, Mandarin, and mixed-language recognition.

Highlights

• Dictate into the current caret with a global Right Alt or F8 shortcut.
• Keep recordings and recognition on the local computer.
• Recognize Mandarin, English, and code-switched speech with SenseVoice Small.
• Restore Chinese/English punctuation without rewriting the transcript.
• Use the guided microphone and shortcut setup on first launch.
• Fall back to the clipboard if focus changes before recognition finishes.

Downloads

• SAID-Setup-0.2.0.exe — per-user installer with the speech models included.
• SAID-windows-x64-0.2.0.zip — portable Windows x64 build.
• SHA256SUMS.txt — checksums for verifying both downloads.

Important: unsigned build

This release is not Authenticode-signed. Microsoft Defender SmartScreen may show an unrecognized app warning. Download SAID only from this GitHub release and verify the file against SHA256SUMS.txt before running it.

SAID supports 64-bit Windows 10 and 11. A practical minimum is a 4-core x64 CPU and 8 GB RAM; 6 or more modern CPU cores and 16 GB RAM are recommended while coding. No dedicated GPU is required.