Skip to content

YazSes v2.4.0-dev.2 — Wave H 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 H — the full 10-feature SoA round from design/vision/v2-research/10-wave-h.md (ADRs adr-v2-055..064). Adds the four features beyond the dev.1 ship-now tier:

New since dev.1 (all off by default)

  • Ambient Audio-Event Guard (ADR-v2-061, [audioguard]) — auto-pauses dictation (or alerts) when it hears a doorbell / phone / alarm / your name. Classifier behind soundawareness.
  • On-Device Condense (ADR-v2-062, [condense]) — a condense variant inserts a tightened summary of your own long dictation instead of the verbatim text.
  • Slot-Filling Dictation (ADR-v2-063, [slotfill]) — one utterance fills a named-field template ("high priority, affects Firefox" → the right fields).
  • Few-Shot Command Spotter (ADR-v2-064, [cmdspotter]) — enroll a few examples of short commands ("send", "stop") that fire instantly without a full decode. Encoder behind kws.

Plus the six dev.1 ship-now features: Emoji & Symbol by Voice, Voice Unit Conversion, Spoken Temporal Normalizer, Mid-Utterance Self-Repair, Spoken Spreadsheet, Clipboard-History by Voice. yazses features now lists 75 capabilities.

Quality

1202 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Every Wave H pure core is unit-tested to 100%; heavy models (PANNs, abstractive LLM, SpeechLLM, KWS encoder) stay lazy behind optional extras. The default install and the v1 dictation path are unchanged.