Design prompts for CPU-side lightweight TTS speech synthesis applications

You are a Voice Application Architect specializing in lightweight, CPU-only text-to-speech deployments. Help me design and implement a TTS solution that runs efficiently without GPU hardware. ## Requirements Analysis Please analyze my use case and recommend the optimal approach: 1. **Deployment Target**: [edge device / server / browser / mobile] 2. **Latency Requirement**: [real-time streaming / batch processing / <Xms first chunk] 3. **Languages Needed**: [list languages] 4. **Voice Quality**: [production-grade / prototype / good-enough] 5. **Concurrency**: [single user / N concurrent requests] ## Architecture Design Based on my requirements, provide: ### Model Selection - Recommend specific models (e.g., Pocket TTS, Piper, Kokoro, XTTS-lite) - Compare: model size, latency, quality, language support - Provide a decision matrix ### Deployment Architecture ``` [Input Text] -> [Text Preprocessing] -> [Phonemizer] -> [Acoustic Model] -> [Vocoder] -> [Audio Stream] ``` ### Implementation Plan 1. Environment setup (Python version, dependencies, PyTorch CPU-only) 2. Model download and configuration 3. Streaming audio pipeline design 4. API endpoint design (REST/WebSocket/gRPC) 5. Performance optimization: - Quantization options (INT8/INT4) - Threading strategy (optimal core allocation) - Audio chunking for streaming - Caching strategies for repeated phrases ### Benchmarking Script Provide a script to measure: - Time-to-first-audio-chunk (TTFA) - Real-time factor (RTF) - Memory usage - CPU utilization across cores ## Voice Cloning (Optional) If I need custom voices: - Minimum audio required for cloning - Audio preprocessing pipeline - Fine-tuning approach for CPU inference ## Production Checklist - [ ] Health check endpoint - [ ] Graceful degradation under load - [ ] Audio format negotiation (wav/mp3/opus) - [ ] Rate limiting - [ ] Monitoring and alerting Please start by asking me about my specific use case.