Skip to content

YazSes v2.3.0-dev.2 — Wave G complete (developer preview)

Pre-release · not published (publish workflows are pinned to v0.x/v1.x; v2 dev tags do not ship). v1.4.1 remains the stable release. Every feature is OFF by default.

Completes Wave G — the full 10-feature SoA round from design/vision/v2-research/09-wave-g.md (ADRs adr-v2-045..054). Adds the six features beyond the dev.1 ship-now tier:

New since dev.1 (all off by default)

  • Compose-in-Target-Language (ADR-v2-049, [compose]) — speak your strongest language, inject in a target language (inverse of Wave D's X→English). Offline MT behind the translate extra.
  • Grammar Repair (ADR-v2-050, [gec]) — minimal-edit article/agreement correction for L2 dictation, register-preserving; built around an is_minimal_edit safety guard.
  • Screen-Grounded Dictation (ADR-v2-051, [screengrounded]) — primes Whisper with names visible on screen so they transcribe right the first time, in any app. OCR VLM behind screenocr.
  • Head-Pointer (ADR-v2-052, [headpointer]) — move + click the cursor by head pose; completes the hands-free bundle. MediaPipe behind headpointer.
  • Silent Lip-Reading (ADR-v2-053, [lipread]) — dictate with no audible voice; a webcam reads your lips. VSR model behind vsr.
  • Sign-Language Input (ADR-v2-054, [sign]) — Deaf/HoH signers dictate by signing to the webcam; ASL recognized on-device. SignGemma behind sign.

Plus the four dev.1 ship-now features: Entity ITN, Redaction Ink, Field-Aware Dictation, Corpus Voiceprint Scrub. yazses features now lists 65 capabilities.

Quality

1153 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Every Wave G pure core is unit-tested to 100%; heavy models (neural ITN, GLiNER PII, kNN-VC, MT, GEC LM, OCR VLM, MediaPipe, VSR, SignGemma) stay lazy behind optional extras. The default install and the v1 dictation path are unchanged.