FR-3: Transcription

Offline transcription of recorded audio to text using Whisper.

Requirement

1. Use whisper-rs to run Whisper (GGUF) models entirely on-device.
2. Default language: English (en). Allow user language selection in Settings.
3. Transcription must complete within ≤ 3 s (95th percentile) for 5-second recordings on Apple
4. Silicon with the small model.
5. Support user-selectable model sizes for latency/accuracy trade-off.
6. No external network calls during transcription.

Rationale

On-device inference preserves privacy and removes network latency, achieving the product’s privacy-first promise.

Acceptance Criteria

• Transcription completes within latency budget on M1 and Intel reference machines.
• Selecting a different model in Settings updates the engine without restart.
• No outbound network traffic observed via packet capture.
• Errors (e.g. model missing) surface in UI overlay/log with actionable message.

Test-Driven Design

Begin with failing automated tests for latency, language selection, and network isolation. Implement transcription until all tests pass, following TDD.

References

PRD §6 Functional Requirements – FR-3