Skip to content

Wave H — SoA research, 10 net-new features

Date: 2026 · Tier: design/ — public engineering research · Author: Mohsen Seyedkazemi Ardebili Companion: fed adr-v2-055 through adr-v2-064. 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 ~52 features existing before this wave (v2 + Waves D/E/F/G). Ranked strongest-first. Anchors web-verified at the time (2025-2026).

  1. Spoken Spreadsheet / Table Mode — 2D cell-addressed dictation + grid nav ("A1 revenue, tab, down, next row"). Anchor: arXiv 0809.3571, Windows Voice Access 2025. Pure: parse_grid_command(text)→KeySequence. Distinct from Field-Aware (form fields, not 2D cells). → adr-v2-059.
  2. Ambient Audio-Event Guard — auto-pause/alert on doorbell/alarm/name/baby-cry. Anchor: SELD wearables (2509.14650), E2PANNs (2506.23437). Pure: label→policy + debounce; PANNs deferred. Distinct from Wake-Word (activation, not environmental awareness). → adr-v2-061.
  3. Spoken Temporal Normalizer — "next Friday at 3" → concrete locale date/time vs the clock. Anchor: SCATE (2507.06450), SUTIME/HeidelTime, dateparser. Pure: resolve_temporal(text,now). Distinct from Entity ITN (formats number words, doesn't resolve relative dates). → adr-v2-057.
  4. Voice Unit / Currency / Number Conversion — "twenty miles in km" → "32.19 km", inline, offline. Anchor: pint registry (an offline Wolfram-Alpha-alternative). Pure: regex grammar
  5. factor table. Distinct from ITN (format) + Spoken Math (render) — this evaluates. → adr-v2-056.
  6. On-Device Condense — a hotkey variant inserts a tightened summary of your own long burst. Anchor: Canary-Qwen-2.5B, Omi 0.6B. Pure: extractive TextRank/lead ranking; abstractive LLM deferred (reuses llm_cleanup). Distinct from Meeting Scribe (others' multi-speaker capture). → adr-v2-062.
  7. Structured Form / Slot-Filling Dictation — one utterance → many named template fields. Anchor: slot-filling for SpeechLLMs (2510.19326, 2510.15851). Pure: schema + keyword matchers → tab-fill; SpeechLLM deferred. Distinct from Field-Aware (formats one field). → adr-v2-063.
  8. Clipboard-History by Voice — "paste the second thing I copied". Pure: ring buffer + ordinal grammar; embedding recall deferred. Distinct from Smart-Paste (single paste formatting). → adr-v2-060.
  9. Mid-Utterance Self-Repair — "email Sarah, no I mean Sara" → "email Sara". Anchor: Interactive Dictation / TERTiUS (2307.04008). Pure: editing-term grammar → span-replace before injection. Distinct from Spoken Edit (post-injection, separate turn). → adr-v2-058.
  10. Emoji & Symbol by Voice — "shrug emoji"→🤷, "right arrow"→→, "degree sign"→°. Pure: name→codepoint table + phrase replace. Distinct from Voice Punctuation (ASCII only). → adr-v2-055.
  11. Few-Shot Personal Command Spotter — low-latency KWS action triggers from few-shot enrollment. Anchor: EdgeSpot (2601.16316), GE2E-KWS (2410.16647). Pure: enroll store + cosine dispatch; KWS encoder deferred. Distinct from Wake-Word (single activation phrase). → adr-v2-064.

Ship-now pure (do first)

1 Spoken Spreadsheet, #3 Temporal Normalizer, #4 Unit Conversion, #9 Emoji & Symbol — fully

dependency-free grammars/lookups extending the existing grammar→dispatch→injector path, each an accessibility win a cloud tool can't offer privately.

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