YazSes v2.6.0-dev.2 — Wave J flagship (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.

Continues Wave J (design/vision/v2-research/12-wave-j.md) with the flagship feature and the runtime formatting toggle:

New since dev.1 (all off by default)¶

• Confidence-Gated Re-Ask (ADR-v2-077, flagship, [reask]) — instead of silently injecting a low-confidence guess, YazSes holds only the uncertain span and interactively resolves it: an A/B pick from a confusable set (their/there) or "say that word again", patching just the placeholder. low_confidence_spans (from the token log-probs faster-whisper already returns) + confusion_set + parse_choice. Distinct from Confidence Ink, which only marks — this repairs and closes the loop. Two independent research streams converged on it.
• Verbatim ⇄ Autoformat Live Toggle (ADR-v2-078, [verbatim]) — say "dictate verbatim" to freeze all ITN / voice-punctuation / reflow (so "one hundred dollars" types literally) and "resume formatting" to restore — mid-burst, no stop. The only runtime ITN on/off switch; essential for code identifiers, quotes, spoken passwords, legal capture.

yazses features now lists 89 capabilities.

Wave J remaining¶

Crowd-Proof Dictation (target-speaker extraction), Self-Learning Correction Dictionary, Voice Fuzzy File Open, Voice Jump-to-Symbol, Recording Import (batch transcription), Spoken Shell Pipeline Builder.

Quality¶

1270 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Both new cores are unit-tested to 100% with no new dependency; the default install and the v1 dictation path are unchanged.