YazSes v2.3.0-dev.1 — Wave G ship-now tier (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.
Opens Wave G — a fresh 2026 SoA round (10 candidate features distinct from the prior 44) in design/vision/v2-research/09-wave-g.md + design/adr/adr-v2-045..048. The four ship-now, dependency-light features:
New features (all off by default)¶
- Entity ITN (ADR-v2-045,
[itn]) — turns spoken structured entities into written form with no command words: "john dot doe at gmail dot com" → the email, "version two point one" → v2.1. Wired on the dictate path. Targets the biggest daily dictation friction. - Redaction Ink (ADR-v2-046,
[redaction]) — detects spoken secrets (card numbers via Luhn, SSNs, API keys, emails, IPs, phones) and masks them before they're typed into another window. Wired on the dictate path;mask/holdmodes. - Field-Aware Dictation (ADR-v2-047,
[fieldaware]) — reshapes output to the focused control: number field → digits, search box → no trailing period, password field → refuses to type. - Corpus Voiceprint Scrub (ADR-v2-048,
[learning] anonymize_audio) — speaker-anonymizes learning-corpus audio before storage, so retained clips keep what was said but not an identifiable voice. Wired in the corpus writer.
yazses features now lists 59 capabilities.
Wave G backlog (medium / hardware-gated)¶
Screen-Grounded Dictation, Compose-in-Target-Language, Silent Lip-Reading, Head-Pointer, Grammar Repair, Sign-Language Input.
Quality¶
1124 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Every Wave G core is unit-tested to 100%; the heavy tiers (neural ITN, GLiNER PII, kNN-VC anonymizer) stay lazy behind optional extras. The default install and the v1 dictation path are unchanged.