feat(vad): implement native voice activity detection for Linux

[wchest]

Summary

Implements native voice activity detection for Linux using CPAL audio capture and the Silero VAD model, providing better platform integration and reliability compared to browser-based audio processing.

Motivation

While the existing web-based VAD works well, browser audio APIs can have limitations on Linux systems. This native implementation provides:

• Platform Integration: Direct OS-level audio capture using CPAL instead of browser MediaRecorder API
• Device Control: OS-level device enumeration and selection without browser restrictions
• Audio Reliability: Native audio streams eliminate browser-specific audio issues
• Local Processing: All audio processing happens locally maintaining the project's privacy-first approach

Implementation Details

Architecture

• Service Layer: New NativeVadService alongside existing VadService with identical interface
• Dynamic Selection: Runtime service selection based on user settings (recording.vad.useNative)
• Rust Backend: CPAL audio capture + Silero VAD processing in src-tauri/src/recorder/vad.rs
• Event Communication: Tauri events for speech start/end with embedded audio data

Key Features

• Configurable sensitivity slider (0.1-0.9 threshold) in recording settings
• Automatic session cleanup to prevent conflicts
• OS-level device enumeration and selection
• Proper state management matching web VAD behavior
• Direct file handling without frontend permission issues

Technical Implementation

• CPAL: Cross-platform audio library for native audio capture
• 16kHz Preference: Optimal sample rate for Silero VAD model
• Event Architecture: Separate speech start/end events for accurate UI state transitions
• Embedded Audio: File contents included in events to avoid Tauri filesystem permission issues
• Platform Gating: Linux-only initially for focused testing and validation

User Experience

Settings Integration

• Native VAD toggle in recording settings (Linux only)
• Real-time sensitivity slider with live value display
• Dynamic description text showing active VAD implementation
• Seamless switching between web and native VAD modes

Testing

• ✅ Toggle native VAD on/off in settings
• ✅ Sensitivity slider functionality across full range
• ✅ Device enumeration and selection
• ✅ End-to-end recording and transcription workflow
• ✅ Session management and cleanup
• ✅ UI state transitions and icon accuracy
• ✅ Error handling and graceful fallbacks

Breaking Changes

None. This is purely additive:

• Existing web VAD remains default and unchanged
• Native VAD is opt-in via settings checkbox
• All existing functionality preserved

Dependencies

Added voice_activity_detector = "0.2.1" to provide Silero VAD model integration.

Files Changed

• src-tauri/src/recorder/vad.rs - New native VAD implementation
• src/lib/services/native-vad.ts - TypeScript service wrapper
• src/lib/settings/settings.ts - Added VAD configuration options
• src/routes/(config)/settings/recording/+page.svelte - UI controls and descriptions
• Various integration files for service selection and query handling

Future Considerations

• Test and potentially expand to other platforms after Linux validation
• Consider additional audio processing configuration options
• Explore integration with other native audio features

---

This implementation maintains Epicenter's local-first philosophy while providing Linux users with improved audio processing reliability through native platform integration.