I Gave My Newsletter a Voice (Literally)

Short summary

Author built a voice chat for their newsletter using LiveKit agents, Azure Speech Services (STT/TTS), and GPT-4, allowing users to ask questions and receive real-time voice responses about newsletter content. Architecture uses hybrid retrieval (vector search + metadata) to handle both semantic queries ('GitHub Copilot') and structural ones ('latest issue'). Streaming speech-to-text and text-to-speech cut perceived latency in half, and the full stack (FastAPI backend, Python agent, Astro frontend) runs in one Railway container with sqlite-vec for vector search—no external database required.

• •Built voice chat using LiveKit agents + Azure Speech Services + GPT-4 for real-time voice conversations
• •Hybrid retrieval (vector search + metadata injection) handles both semantic and structural queries over newsletter content
• •Streaming architecture reduces latency; entire stack runs in one Railway container with sqlite-vec for vector search

Generated with AI, which can make mistakes.