Build a local voice assistant that keeps speech, reasoning, and synthesis offline while preserving practical latency on consumer CPU hardware.

• Zero internet dependency.
• Hindi-first interaction loop.
• Consumer CPU latency target below one second.
• Modular plugin architecture without changing the core inference loop.

• openWakeWord detection triggers the pipeline.
• Whisper-small performs local ASR.
• Qwen2.5-3B-Instruct Q4_K_M handles local reasoning with 128K context.
• Piper TTS and fine-tuned XTTS v2 produce speech output.
• Eight-layer plugin architecture isolates tools, memory, routing, and generation.

• ASR and TTS latency under CPU-only constraints.
• Hindi corpus quality for voice persona fine-tuning.
• Context management with local quantized model memory.
• Tool plugin boundaries in an offline runtime.

• Use quantized local model execution.
• Fine-tune XTTS v2 on AI4Bharat Hindi corpus.
• Keep plugin interfaces thin and deterministic.
• Optimize each stage independently before end-to-end latency tuning.

• Measured wake-to-response end-to-end latency.
• Profiled ASR, LLM, and TTS stages separately.
• Tested offline operation with no network dependency.
• Validated new plugin integration without modifying core runtime.

• Achieved under 800ms end-to-end latency on consumer CPU.
• Kept the entire voice pipeline offline.
• Supported modular plugin extension with stable core interfaces.

• Offline assistants are latency orchestration problems as much as model problems.
• Language-first UX changes tokenizer, ASR, TTS, and memory decisions.
• Local privacy constraints make deterministic system boundaries valuable.