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.