Add on-device audio transcription with speaker diarization

[ldt]

Summary

Adds automatic post-recording transcription with speaker diarization to TabRecord, using WhisperKit and SpeakerKit Swift packages. All processing runs on-device using Apple Neural Engine acceleration — no audio is sent to external servers.

Key Changes

• TranscriptionCoordinator: Central orchestrator that manages the full pipeline (model download → audio prep → ASR → diarization → alignment → file output) and reports progress to the UI
• ModelManager: Actor that handles CoreML model download and caching for WhisperKit and SpeakerKit models in ~/Library/Application Support/TabRecord/Models/
• AudioPreprocessor: Extracts the selected audio track from multi-track MP4 recordings and resamples to 16 kHz mono PCM (Whisper's required format)
• WhisperKitTranscriber: Thin wrapper around WhisperKit SPM that produces word-timestamped segments with language detection
• SpeakerKitDiarizer: Thin wrapper around SpeakerKit SPM that produces speaker-labeled time segments using pyannote v4 models
• SegmentAligner: Merges word-level transcription with diarization segments to produce speaker-attributed transcript
• TranscriptWriter: Serializes aligned segments to .txt, .json, and .srt output formats
• TranscriptionPreferences: @AppStorage-backed model for user preferences (model variant, language, speaker count, audio source, output formats)
• PreferencesView: SwiftUI preferences panel with Transcription tab (macOS 14+)
• MenuBarController: Updated to trigger transcription after recording stops and display transcription progress in the menu
• AppDelegate: Updated to request notification permissions for transcription completion/failure alerts

Implementation Details

• Hardware utilization: Uses cpuAndNeuralEngine compute units to maximize ANE throughput on Apple Silicon (encoder on ANE, decoder on CPU)
• Privacy: Fully on-device processing; models cached locally after first download (~1.5 GB for large-v3-turbo)
• Background processing: Pipeline runs at .utility priority on a detached Task to keep UI responsive
• Error handling: Graceful failure with user notifications; recording files are preserved if transcription fails
• Minimum macOS: Requires macOS 14.0+ for transcription (recording still works on 13.0+)
• Output: Transcript files saved alongside recordings with speaker labels and timestamps

Testing

Comprehensive unit tests added for:

• ModelManager (cache hits, disk space checks, model listing)
• AudioPreprocessor (track extraction, resampling, temp file cleanup)
• SegmentAligner (empty inputs, single/multi-speaker alignment, boundary cases)
• TranscriptWriter (TXT/JSON/SRT formatting, timestamp conversion)

https://claude.ai/code/session_01DkKXRRkUErWxE49oAvcxwE