Productionize Live Notes with on-device Whisper transcription

[ucguy4u]

Problem

Live Notes is now merged into Assistant at /agent/live-notes and Pixel 9 has a working Android SpeechRecognizer bridge with offline preference. That is a useful smoke-tested bridge, but it is not the production-grade on-device Whisper path we want for long team conversations, privacy guarantees, diarization-ready chunks, and consistent offline behavior.

Scope

In scope:

• Replace or augment the Android SpeechRecognizer bridge with a real on-device Whisper/whisper.cpp runtime behind com.airo.meeting/transcription.
• Add iOS parity behind the same Flutter MeetingTranscriptionService contract.
• Preserve explicit-start listening only.
• Keep raw audio and transcript local by default.
• Emit partial and final transcript chunks with stable timestamps.
• Preserve pause, resume, stop, and permission-denied error behavior.
• Validate on Pixel 9 and at least one iOS simulator/device path.

Out of scope:

• Cloud transcription fallback.
• Silent/background always-on listening.
• Uploading raw audio or raw transcript text.
• Full task/reminder creation workflow beyond surfacing extracted action candidates.

Acceptance criteria

• Live Notes uses a real on-device Whisper path on Android when the model is available.
• Missing model, denied microphone permission, low storage, and native failure states are explicit in UI/state.
• Pixel 9 test confirms real spoken audio produces partial/final transcript chunks.
• Pause/resume stops and restarts transcript emission without losing final chunks.
• Stop finalizes the latest pending text into local intelligence.
• No network upload path exists for meeting audio or transcript text.
• Focused meeting tests pass.
• flutter analyze is clean for the touched surface.
• Android debug build installs and launches on Pixel 9.

Agent

Primary: agent/ai-llm
Supporting: agent/mobile-ui

Estimate

3-5 days depending on model packaging/download strategy.

Risks

• Battery drain during long conversations.
• Model size/storage pressure on mid-range devices.
• Native audio lifecycle bugs around interruption, Bluetooth microphones, and app backgrounding.
• Privacy regressions if fallback paths accidentally upload audio or transcript text.

Context

Merged baseline: ccec84f feat(agent): merge assistant model setup
Current route: /agent/live-notes
Compatibility redirects: /meetings, /agent/meetings
Current Android bridge: android.speech.SpeechRecognizer with EXTRA_PREFER_OFFLINE.

[ucguy4u]

Required Before Implementation: Use Cases + Automation Flow

Before coding this ticket, add a deterministic implementation note or child task covering:

• Primary user journey: actor, preconditions, trigger, happy path, success state.
• Alternate/failure paths: offline, missing model, permission denied, runtime/tool failure, cancel/retry, background/foreground.
• Automation flow: Given/When/Then steps, fixtures, mocks/stubs, UI assertions, state/database assertions, notification/tool-trace/memory assertions where relevant.
• Hardening checks: privacy redaction, accessibility, timeout/performance bounds, cleanup/reset.
• Delegation prompt: Ask a QA/test agent to write the automation flow before implementation begins.

Tracker: #323