🎙️ Streaming Speech-to-Speech

Record your voice → stop recording → hear the AI respond with streaming audio.

Pipeline: Moonshine Tiny (ASR) → Qwen2.5-0.5B (LLM, CPU) → Kokoro (TTS)

Based on 20 experiments optimizing from 1000ms → 225ms. Demo uses CPU LLM for reliability.

How it works: Record → Moonshine transcribes (CPU) → Qwen generates response (CPU) → Kokoro synthesizes streaming audio (CPU). LLM runs on CPU for ZeroGPU reliability.

View benchmarks & code