Documentation reconciliation brief
Scope: edit only docs/guides/transcription-adapters.md and
docs/guides/transcription-adapter-testing.md in this Hudson worktree.
Do not modify code, manifests, the proposal, progress report, or other worktrees.
Do not commit, push, publish, install, download models, or run provider calls.
Replace stale future-tense and pseudocode-only sections with a usable account of
current source. Read the actual public types and adapters under
packages/native/apple/HudsonKit/Sources/HudsonTranscription* and tests beside
these modules. Keep shared core independent of vendor SDKs and TalkieKit.
Show actual registration, configuration, model selection, readiness/preparation,
file submission, caller-fed PCM lifecycle, event consumption, cancellation,
terminal outcomes, normalized results/provenance, and extension instructions.
Use compiling-shape API names from source, not invented pseudocode names.
Describe remote ElevenLabs and local WhisperKit as reference implementations.
WhisperKit is batch only; Parakeet stays preferred. Model additions use catalog
metadata and compatibility rather than generic UI branches. Do not claim a
runtime-updated catalog if the source uses a static catalog.
Read host integration (read-only) from
/Users/arach/dev/talkie-worktrees/transcription-adapters/apps/macos/TalkieTranscription,
main app Services/Transcription, and TalkieAgent Services/SelectedTranscriptionService.swift.
Global AgentController now fixes its binding at capture start, receives PCM only
after recorder file writes, streams to live-capable models, and uses finalized
files for batch-only models. Live failure must not fall back to batch or legacy.
Do not claim real-device acceptance or installed app replacement.
Testing guide must distinguish fixture coverage, native SDK compilation, app integration test/build coverage, and live model/vendor acceptance. The progress report is chronological and includes superseded entries: read latest entries and current logs rather than copying old counts. As of this request: 14 WhisperKit standalone fixtures passed, host suite reports 18 tests passed; agent suite latest completed run reports 5 tests with one parameterized test covering two cases, and an additional cancellation test is still running. Main app validation build passed using an isolated Termini packaging fix; canonical published Termini binary was not changed. Do not describe the published dependency path as accepted.
Live MAI/Gemini/ElevenLabs network behavior and real-model WhisperKit quality, latency remain unverified. No usable WhisperKit installation was found in the normal Talkie model locations; do not generalize that to the entire machine. MAI-Transcribe-2, dedicated Gemini 3.5 file/live, and conversational Gemini 3.8 Live input transcription are separate adapters/paths; don't blur them.
Deliver a concise completion report with changed files, source points verified, and remaining factual uncertainties. Primary agent will review both guides.
Latest acceptance update: the progress report now includes a passing native Parakeet V3 test for file and caller-fed live transcription using installed assets and generated non-private speech. Read that entry; native quality and latency coverage is limited to that one sentence. The latest agent run passed six tests with three live-routing parameter cases. Global microphone acceptance and remote vendor acceptance are still unverified.