Voice‑Driven Murder Mystery Game Lets Users Interview AI Suspects

A developer has unveiled a voice‑AI interrogation platform that leverages OpenAI’s GPT‑Realtime‑2.1 via WebRTC to conduct speech‑to‑speech conversations. The system authenticates users through Clerk and imposes a 30‑minute session limit to control the high cost of the model. During an interrogation, each “suspect” can be called upon to provide evidence. The user’s accusations are recorded, and a separate GPT‑5‑mini judge evaluates the evidence presented, accepting paraphrased statements but rejecting vague or fishing‑type claims.

The project, originally prototyped two to three years ago when voice AI was still nascent, has been rebuilt with modern tools such as Next.js, MongoDB, and Clerk for authentication. The developer notes that the system’s cost constraints required the introduction of user‑based restrictions and a session timer. The platform’s design also includes a “tool” that captures the suspect’s response and the evidence list supplied by the interrogator, which the judge then cross‑checks against the case’s required facts.

While the project is still in early stages, it demonstrates a novel approach to combining real‑time voice AI with automated evidence verification. The developer invites the community to test whether the suspects hold up under a realistic interrogation scenario, and has shared a link to the discussion thread on Hacker News.