Mirror the personal dictionary through iCloud Key-Value Store with
deterministic merge rules, main-app-only sync UI, and App Group as the
keyboard runtime cache. Add cloud-engine ASR with dictionary bias and
expand the bundled custom language model lexicon.
Reduce perceived latency from key release to final text:
- Adaptive chunking: 2.5s first chunk + 5s follow-ups so short
utterances start on-device recognition while still recording.
- Session-level ASR warmup and audio-format cache reuse to remove
per-utterance cold-start of SpeechAnalyzer.
- Mirror live pipelined partials to the keyboard transcript line via
a new flow.transcriptionPartial App Group key + Darwin ping.
Also commits the accumulated custom language model, Flow session,
keyboard extension restructure, and Xiaomi MiMo provider work in
progress on this branch.
Build a first-version SFCustomLanguageModelData for on-device ASR
customization, generated offline on macOS (no app code dependency).
- Scripts/lexicon: Python builders for Sogou-derived phrases (scel +
SogouPopularDict) and a curated bilingual AI/tech/brand seed lexicon
- export_clm.swift / prepare_clm.swift: macOS CLI tools that export the
.bin training asset and compile it into LM + Vocab via the Speech
framework
- Resources/CustomLanguageModel: generated phrases.tsv, manifests, the
129k-phrase .bin, and compiled LM/Vocab assets (zh_CN)
- .gitignore: ignore lexicon build cache and Python bytecode
Note: Sogou-derived data is for internal experimentation only; the
curated AI/tech seed is MIT and safe to ship.
Rocky