[ANT1OFF]

Dictation now spells the names of the project you're working in the way the code does, plus your own word list. It's layer 1 of the plan, a pass after the commit. Layer 2, sherpa-onnx hotwords on Parakeet, was measured and not kept (see below).

What changes

• Spelling after commit: apps/web/src/dictation/vocabulary.ts is pure. It runs on each committed utterance, after feat(voice): spoken symbols, formatting and filler clean-up #99's postprocess() (committedText → applyVocabulary), never on the live partial.
  • Case and spacing variants: "use dictation" → useDictation, "dictation dot TS" or dictation.ts → dictation.ts, "feat slash voice vocab" → feat/voice-vocab.
  • The user's "sounds like" spellings are replaced exactly.
  • A conservative near-miss match for long names ("Nematron" → Nemotron, "feet slash voice vocad" → feat/voice-vocab).
• Precision rules:
  • A run made only of common words ("set state", "list file names", "plan clip") is never turned into a project name unless "dot" or "slash" is said in it.
  • A single common word is never replaced unless the user lists it as a spelling.
  • Two names of the same kind that sound alike cancel out. Symbols beat file stems, which beat folders, branches and titles.
  • A near miss is taken only for names of 8 or more letters: within 1 edit, or 2 at 12 letters and over. It must be the unique nearest, and never when the difference is an ending ("Parakeets" stays plural).
• Project vocabulary (fs.vocabulary RPC, packages/server/src/fs/vocabulary.ts):
  • Sources: tracked file names and stems (git ls-files), folder names, top-level exports found by a capped git grep -o over TS/JS/Rust/Python, local branches, and words from the project's recent thread titles.
  • Only distinctive names are kept (distinctiveTerm).
  • Files and symbols from the last 200 commits come first.
  • It is fetched only once dictation is prewarmed or started, never on opening a thread.
• Settings, Voice page → "Vocabulary" group: a "Use this project's names" switch (on by default) and "Your words", a list with "sounds like" spellings that saves on each change. The "Vocabulary hints" copy no longer says hints are unsupported.

Limits

• Caps: 3,000 project terms in all, with at most 1,500 symbols, 1,200 file names, 300 folders, 150 branches and 150 title words. Up to 200 thread titles are read. The user's list holds 500 words with 8 spellings each.
• Git work: each git call is bounded at 5 s; the export grep's output is capped at 2 MB.
• Cache: kept per checkout and read again when HEAD or the index's mtime changes, or after 10 minutes. 8 checkouts are cached at once.
• Locality: everything stays local. Nothing logs a term or transcript text.

Hotwords: measured, not kept

Clips: Parakeet TDT v2 int8 offline, through sherpa-onnx-node 2.0.0, with 2 threads like the product. 12 synthetic say clips × 2 voices (Samantha, Daniel): 10 name repo identifiers and 2 are controls of only common words.

Hotword setup: they only apply with modelingUnit: "bpe" and a bpe.vocab, which these exports don't ship. One was generated from tokens.txt with scores of −id. Pre-tokenized phrases are rejected ("Encode hotwords failed").

Metrics:

• WER is over the spoken words, with case and separators normalised.
• "Names" counts target identifiers that appear verbatim in the final text.
• "False" counts control clips that layer 1 changed.

Condition	WER raw	WER + layer 1	Names raw	Names + layer 1	False on controls
greedy (today)	10.9 %	6.1 %	4/24	22/24	0
modified beam search, no hotwords	10.4 %	6.1 %	6/24	22/24	0
beam + hotwords (2,835 project terms), score 0.5	6.5 %	5.7 %	8/24	22/24	0
beam + hotwords (2,835), score 1.0	6.1 %	5.2 %	8/24	22/24	0
beam + hotwords (2,835), score 1.5	8.3 %	7.8 %	8/24	21/24	0
beam + hotwords (the 12 targets only), score 1.0	5.7 %	5.7 %	8/24	22/24	0

Why hotwords weren't kept:

• Most of the gain is already layer 1's. Layer 1 alone takes names from 4/24 to 22/24 and WER from 10.9 % to 6.1 %.
• Hotwords add about two words in roughly 230 at best (6.1 % → 5.2 %), on TTS audio, with no gain in names.
• They are score-sensitive. At 1.5 they cut the end off an utterance in both voices ("voice input section dot tsx" → "voice input sects"). A lost word is worse than a misspelling.
• They change today's decoding. Accurate mode would have to switch from greedy to beam search, with a vocab file whose scores are invented.
• Latency wasn't measured reliably: the machine's load average was about 30 while the bench ran (200–565 ms per clip, all conditions overlapping).

The two names layer 1 still misses:

• "inaccurate.ts" for accurate.ts: "inaccurate" is a real word.
• "Sherponx" / "sherpa onks" for sherpa-onnx: 2 edits on a 10-letter name.

Merge with #98 and #99

• feat(voice): spoken symbols, formatting and filler clean-up #99's clean-up runs first, so "dot", "slash" and "underscore" are already characters when names are spelled. The matcher handles both spoken and glued forms. A unit test covers the chain.
• Main is broken by those two PRs together. feat(voice): spoken symbols, formatting and filler clean-up #99's e2e test "what commits is cleaned up…" still opened Voice through Advanced, which feat(voice): a normal Voice setting with a push-to-talk key picker #98 moved to the top-level nav, so it timed out on main. This PR switches that one line to the Voice nav link.

Checks

• bun run typecheck, bun run lint, bun run fmt:check, bun run check:scaffold, bun run check:docs: pass.
• bun run test over dictation, speech process and module, server fs and the fs handler, contracts rpc, settings and vocabulary, and web settings and composer: 97 files, 979 passed, 1 skipped.
• playwright test e2e/dictation.spec.ts: 21 passed. That includes the new test: a committed utterance spells retry.test.ts and the user's "para keet" → Parakeet, and with "Use this project's names" off only the user's word is spelled.
• No full bun run check of my own: the machine is shared.

Known limits:

• Synthetic success only: no live microphone test.
• Project names apply only in a thread's composer. The /new and /new-task drafts get the user's list alone.

Screenshots

Settings → Voice input with a word list, the whole page (dark)

Before	After

Settings → Voice input with a word list, the whole page (light)

Before	After