Skip to content

Wave I — SoA research, 10 net-new features

Date: 2026 · Tier: design/ — public engineering research · Author: Mohsen Seyedkazemi Ardebili Companion: fed adr-v2-065 through adr-v2-074. See the waves index.

A snapshot of the field, kept as the research record behind the ADRs it fed. Feature status should be checked against yazses features and the linked ADRs, not this note.

All on-device, off by default, distinct from the ~62 features existing before this wave (v2 + Waves D-H). Ranked strongest-first. Anchors web-verified at the time (2025-2026); items #5-#10 rest on engineering substrates (AT-SPI2, WM IPC, CSL) rather than a single paper — flagged honestly in the original sweep.

  1. Hard Contextual Biasing — decode-time trie/WFST hotword boosting so rare names/jargon actually win (vs soft initial_prompt). Anchor: WCTC-Biasing (2506.01263), trie K-step (2509.09196), sherpa-onnx hotwords. Pure: build_hotword_trie + bias_logits/rescorer. Distinct from Context-Primed (soft prompt) + Personal Adapter (LoRA). → adr-v2-069.
  2. Diarized Conversation Capture + Rename-by-Voice — attributed per-turn Markdown, rename "speaker two" → "Alice". Anchor: pyannote community-1 / pyannote.audio 4.0. Pure: SpeakerLabelMap + apply_rename + render_attributed_markdown. Distinct from Meeting Scribe (bulk) + Multi-User Profiles (auth). → adr-v2-074.
  3. Adaptive Latency Governor — load-aware decode policy + speculative decoding (a distil draft model matching identical output ~2× faster). Anchor: HF spec-decode, distil-large-v3.5 (2311.00430). Pure: pick_policy(load, config). Distinct from Ghost-Ahead (pre-warm only). → adr-v2-073.
  4. Per-Language Auto Model Switching — LID first ~1-2s → hot-swap a language-specialized model. Anchor: Whisper LID + per-language distil/LoRA. Pure: route_language(lid, registry). Distinct from Code-Switch (within-utterance) + Polyglot (fixed pair). → adr-v2-072.
  5. Voice-Driven Document-Wide Find-and-Replace — "replace every utilise with use". Anchor: AT-SPI2 EditableText. Pure: parse_replace_command + to_editor_actions. Distinct from Spoken Edit (last utterance only). → adr-v2-068.
  6. Terminal Command Safety Gate — hold rm -rf/curl|sh until spoken confirm when focus is a terminal. Pure: classify_command_risk ruleset + confirm state machine. Inspects command semantics before injection. → adr-v2-065.
  7. Structured-Markup Dictation — speak Markdown/org tables and lists. Pure: render_markup + structure grammar. Distinct from Spreadsheet (drives an app) + Code/Math. → adr-v2-067.
  8. Voice Window/Workspace Management — "move window left half", "workspace 3". Anchor: wmctrl/hyprctl/swaymsg. Pure: parse_wm_command + backend adapters. Distinct from Mouse Grid/Pilot. → adr-v2-070.
  9. Spoken Regex Builder — "four digits dash two digits" → \d{4}-\d{2}. Pure: nl_to_regex compositional grammar. Builds search patterns, not literal code. → adr-v2-066.
  10. Citation-by-Voice from local BibTeX/CSL — "cite Vaswani 2017" → formatted citation, offline. Pure: resolve_citation fuzzy match + format_citation. Distinct from RAG (retrieve+generate). → adr-v2-071.

Ship-now pure (do first)

7 Structured-Markup, #6 Terminal Command Safety, #5 Find-and-Replace, #9 Spoken Regex —

cleanest fully dependency-free cores (pure text functions), each an accessibility/safety win a cloud tool can't offer.

Citations here have not been re-verified against research/verify_refs.py.