[real tutorial]: a ChatGPT voice app is a two-stage pipeline, not one call. Stage 1: send the conversation to the LLM and get the text response. Stage 2: send that text to Murf's TTS API for the voiceover. Three things the tutorial gets right: 1. Keep API keys in a .env file (VITE_OPENAI_API_KEY style), never in the source you share. 2. Maintain conversation context across turns; the voice is only as good as the response it is reading. 3. Design for sequential latency. You cannot synthesize speech before the LLM finishes writing, so the UX needs to account for LLM time plus TTS time. Streaming the TTS on partial sentences (see the Falcon streaming pattern) is the upgrade path. At the time this was written Murf API access was request-based; check the current dashboard for self-serve keys.

Context: Web tutorial (dev.to, Joel Gee Roy, Mar 2023): building a voice app that chains ChatGPT and Murf. Flow: user prompt -> ChatGPT generates the response -> the response text goes to Murf's TTS API for the voiceover. Built with React, Vite, TypeScript, the openai npm package and axios, with API keys in a .env file. The key integration lesson: the pipeline is sequential by nature (you cannot speak a response that does not exist yet), so design the UX around that latency instead of pretending the voice is instant.