Kokoro-tts

[TerminallyLazy]

• Added Kokoro TTS support in preload and run_ui scripts.
• Introduced a new API endpoint for text-to-speech synthesis.
• Updated settings to include TTS enable/disable option.
• Refactored speech handling to utilize a centralized speech store.
• Enhanced UI with new speech button and SVG icon.
• Updated dependencies in requirements.txt for Kokoro TTS.

Docker Image Size Impact

• Current: ~9.17 GB
• Addition: ~400-450 MB (4-5% increase)
  • Kokoro model: ~350 MB (or 80 MB if quantized)
  • Dependencies: ~50-100 MB

Memory Consumption Impact

• Base memory: +80-150 MB when model is loaded
• Active TTS: +200-400 MB during audio generation
• Peak usage: +500-600 MB during heavy TTS workloads

Key Optimizations Already in Place

1. Lazy loading - model only loads on first TTS request
2. Single instance - shared pipeline across requests
3. Async processing - non-blocking operations
4. Text chunking - 300-char limits for memory efficiency

Recommendations for Further Optimization

1. Model quantization - reduce from 350MB to 80MB (75% savings)
2. Memory monitoring - track actual usage patterns
3. Resource limits - set container memory caps for TTS processes

The impact is quite modest - about 5% image size increase and reasonable memory overhead. The Kokoro model is specifically designed for edge devices, so it's already optimized for low resource usage.