이미지: METAL LAB 생성
Summary
- Developer MrRowTheBoat posted a web-based mystery game where players interrogate suspects by voice on Hacker News, earning over 200 upvotes
- Conversations run on OpenAI's gpt-realtime-2.1 model via WebRTC, using a voice-to-voice format
- When a player accuses a suspect, the gpt-5-mini model acts as a separate judge, only accepting evidence that was actually presented
- 개발자
- MrRowTheBoat (해커뉴스 사용자)
- 음성 모델
- OpenAI gpt-realtime-2.1, WebRTC 기반 음성-음성 대화
- 판정 모델
- gpt-5-mini가 진술된 증거의 진위를 별도로 심사
- 이용 제한
- Clerk 인증 계정 필요, 세션당 30분 타이머 적용
- 기술 스택
- Next.js, MongoDB, Clerk
- 반응
- 해커뉴스 게시 후 추천 200개 이상 획득
A murder case interrogated by voice
A mystery game that lets players directly question suspects and catch them in inconsistencies to name the culprit has gone viral on Hacker News. The project, released by developer MrRowTheBoat, stands out for letting players interrogate suspects using actual voice rather than text. Conversations run on OpenAI's voice model gpt-realtime-2.1 combined with WebRTC in a voice-to-voice format. The developer noted that model costs are significant, and has restricted access to Clerk-authenticated accounts with a 30-minute limit per session. When a player makes a final accusation, a separate tool is called to log who was accused and what evidence was actually cited during the conversation. The gpt-5-mini model then reviews this record to judge whether the evidence genuinely required to solve the case was actually presented. Rephrasing is accepted, but vague suspicion or leading questions are not counted as evidence, according to the developer.
Why this is possible now
The project originally started as a prototype built two to three years ago. Even then there was an AI boom, but voice AI agent technology was still in its early stages, and the developer reportedly wasn't satisfied with the result. The recent return to the project appears to be driven largely by how much voice AI technology itself has advanced. Real-time voice conversation systems have long been considered a difficult problem, requiring "turn detection" to determine when a person has finished speaking, managing response latency, and maintaining seamless continuity between utterances during a conversation—all at once. OpenAI, in unveiling its third-generation voice system architecture GPT-Live on August 4, said it adopted a design where a full-duplex voice model replaces a separate turn detector, while offloading tasks requiring deep reasoning asynchronously to GPT-5.5. The growing number of projects like this game that use voice conversation itself as a core game mechanic can be seen as evidence that this underlying technology is rapidly stabilizing. The design choice to hand off evidence judgment to a separate scoring model is also notable. When a single model handles both conversation and judgment, consistency tends to suffer—so the roles were split, with a real-time voice model handling dialogue and a separate language model handling rule-based judgment.
So what does this change
The game itself is less a commercial service than a hobby project by an individual developer. But it demonstrates that interactive content built on voice AI can now be made even at the level of a solo developer. Just a few years ago, building a voice agent required combining separate speech recognition and synthesis engines, but now a single real-time voice model paired with a WebRTC connection is enough to create a conversational character. At the same time, cost remains an obstacle. The developer's session time limits and login authentication likely reflect the fact that calling real-time voice models is far more expensive than using text-based models. Given these constraints, voice AI-based interactive content is likely to first proliferate in the form of small-scale experiments for the time being.
