Skip to content

YazSes v2.6.0-dev.5 — Wave J 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 J — the full 10-feature SoA round from design/vision/v2-research/12-wave-j.md (ADRs adr-v2-075..084). Adds the final two pure cores:

New since dev.4 (all off by default)

  • Recording Import (ADR-v2-083, [recimport]) — batch-transcribe existing voice memos, lectures and meetings offline to .txt/.srt/.vtt with timestamps; your audio archive becomes searchable without the cloud. merge_word_timestamps (gap/length segmentation) + write_srt / write_vtt / format_timestamp. Distinct from live Meeting Scribe/Diarize (pre-recorded batch). The Parakeet/NeMo backend is deferred (faster-whisper is the zero-extra fallback).
  • Crowd-Proof Dictation (ADR-v2-084, experimental, [crowdproof]) — continuously reconstruct your enrolled voice out of overlapping babble before STT, so dictation survives an open-plan office or café. The pure numpy plumbing — frame_signal, apply_target_mask, overlap_add (Hann-windowed, normalized) — ships now; the Conv-TasNet/SSM TSE model is deferred behind a crowdproof extra. A different mechanism than the Cocktail Filter (continuous reconstruction vs sub-second gating).

Wave J — all 10 features

Phonetic Spelling, Voice Git Choreographer, Confidence-Gated Re-Ask (flagship), Verbatim⇄Autoformat Toggle, Self-Learning Correction Dictionary, Voice Fuzzy File Open, Voice Jump-to-Symbol, Spoken Shell Pipeline Builder, Recording Import, Crowd-Proof Dictation. yazses features now lists 95 capabilities.

Quality

1295 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Every Wave J pure core is unit-tested to 100%; heavy models (Parakeet, TSE, the confidence head, EmbeddingGemma, NL2Bash SLM, LSP feed) stay lazy behind optional extras. The default install and the v1 dictation path are unchanged.