YazSes v2.0.0-dev.2 — Voice-First Interaction Layer (developer preview 2)¶
Pre-release · not published to PyPI/Snap/APT (publish workflows are pinned to the v0.x/v1.x lines; v2 dev tags intentionally do not ship). v1.4.1 remains the stable release. Every v2 feature is OFF by default.
dev.2 completes Wave A (4) + Wave B (5) of the v2 line — all nine flagship features are now implemented, tested, and integrated (heavy ones behind opt-in extras). Still 100% on-device and privacy-first (ADR-011).
Wave A — fully wired into the daemon¶
- Confidence Ink (ADR-v2-001) — flags low-confidence words from Whisper's token probabilities; wired into the decode path (metadata-count only, no transcript stored).
- Prosody pause→sentence (ADR-v2-002) — a sentence-length pause inserts a period.
- Spoken Edit Mode (ADR-v2-003) — voice edits of the last dictation ("change X to Y"); command-key gated, reuses the Punch-In delete-and-retype.
- Context-Primed Dictation (ADR-v2-004) — transient active-window/selection/clipboard terms fold into the STT prompt; never stored.
Wave B — new packages (off by default)¶
- Personal Adapter P1 (ADR-v2-009) — corpus n-gram + unigram prompt-mining (
personalize/) biases STT toward your jargon; reads the encrypted corpus once, cached and guarded. - Spoken Recall & Ambient Scratch (ADR-v2-005) —
recall/package: search past dictations (yazses recall <query>) and capture spoken notes-to-self (yazses scratch). Corpus-local; nothing leaves the machine. - True Code-Switch (ADR-v2-008) —
polyglot/router.pyrouting layer; dormant until you supply an out-of-band code-switch adapter. - Voice-to-Tool / Spoken MCP (ADR-v2-006) —
agent/planner + confirm guard (all|writes|none); SLM + MCP client behind theagentextra. - AT-SPI Voice Pilot (ADR-v2-007) —
pilot/command parser + label matcher for "click Save"-style desktop control; pyatspi backend is lazy/system-package, labels only, no screenshots.
Manage all of these with yazses features (24 capabilities registered, all v2 ones off by default). New config sections: [confidence] [context] [recall] [agent] [pilot] plus commands.spoken_edit and prosody.pause_sentence_ms.
Quality¶
850 tests green across Linux × macOS × Windows on Python 3.11 and 3.12.
What's next (toward v2.0.0)¶
Wave A UX follow-ups (Confidence overlay markers + voice re-pick; Spoken Edit destructive-confirm), then the Wave C experimental set (Gaze-Routed, sEMG/Modality Router, Accessibility Continuum, Glasses↔Desktop bridge). v2.0.0 ships to all channels once the publish workflows are extended to the v2.x line.