Worked example: when voice adds value over chat in a self-hosted interview workflow

Here is a hypothetical example. A senior backend candidate is asked to diagnose a queue that occasionally processes jobs twice. In chat, they can carefully structure assumptions, provide pseudocode, and revise their answer. That is useful when the goal is evaluating technical precision. In a voice workflow, the same prompt can test a different skill. The interviewer can ask an adaptive follow-up like, “How would your approach change if consumers crash after committing the side effect?” The candidate has to explain the tradeoff conversationally. Voice is not universally better, though. Transcription can lose technical terms, audio is more sensitive to retain, and some candidates communicate more clearly in writing. A practical setup could use chat for detailed design work and voice only where spoken explanation matters. I built Aural to experiment with the voice side of that workflow. It is MIT licensed, 100% TypeScript, and self-hostable with Next.js, Supabase, and PostgreSQL. To start, clone the repo, follow the documented environment setup, connect a supported or OpenAI-compatible LLM API, and keep recordings, transcripts, and scoring data on infrastructure you control with Row-Level Security. I’ll reply to the asimovs-auditor comment with the required AI disclosure. Where would you choose voice over chat, if anywhere?