Vaudeville

My first instinct when I loaded Vaudeville was to treat it the way I treat an experimental Paradox feature branch: fascinating in ambition, potentially unstable in execution, and worth understanding on its own terms before passing judgment. The pitch is genuinely interesting. You step into the shoes of Detective Martini, a private investigator dropped into a moody European city to solve three murders, and every conversation with every suspect is generated by a live AI in real time. You type or speak into a microphone, the nine suspects respond in character, and the interrogation goes wherever the exchange takes it. No branching dialogue tree, no pre-written options. Just you, your questions, and whatever the AI decides your witnesses know. In practice, the depth of that system depends entirely on how sharp your interrogation instincts are. Characters like the arrogant Count Gravesen or the enigmatic lion tamer Monsieur Saxabar each carry their own personality parameters, and the AI does a reasonable job of holding personas across a conversation. A useful in-game notepad sits at the side of the screen, letting you jot down leads to chase and contradictions to press. The investigative loop has real pull when momentum builds. The problem is getting that momentum started. There is almost no tutorial, no starter clue to anchor your first conversation, and characters will simply clam up or stonewalled you if you rub them the wrong way early. One reviewer noted they managed to permanently anger Chief Gretzky, the entry-point contact, within the first few minutes and lost access to key information. For a game where every data point is verbal and self-directed, that kind of dead end hits harder than it would in a game with visible evidence rooms or a case file to fall back on. Technical friction is a real factor in the current build. AI-generated voices are robotic and prone to overlapping when the engine stalls, dialogue can hang for five to fifteen seconds while the system computes a response, and character animations during conversations amount to NPCs standing still and staring past you. The Studio acknowledges this openly, noting that running LLMs locally is GPU-demanding and that strange character behaviour and longer load times are expected while optimisation continues. The good news is that Bumblebee clearly intends to keep building. Post-launch they shipped a reworked game flow with modular world-building tools, plus the Vaudeville Story Vault and a built-in editor accessible from the Workshop button in the main menu. That means user-generated stories and official expansions are already in the pipeline, which matters a lot for long-term value. Steam reviews sit at a Mixed rating, roughly 49 percent positive from around 271 users, which is an accurate read on the split between players who find the AI chaos charming and players who hit the loading pauses and walk away. Who is this actually for? Not players who want the tight, scripted satisfaction of a traditional whodunit. This is for people who are curious about what interrogation feels like when the suspect can actually surprise you, who are patient enough to treat a bugged, experimental early-stage product as a curiosity rather than a finished experience, and who have some tolerance for the uncanny-valley texture of AI voice delivery. The Steam Workshop support means the content ceiling is open-ended if the community engages. That is a real differentiator. At this stage, Vaudeville is more proof-of-concept than polished game, but if the concept genuinely excites you, there is something here worth poking at. Diego, Scout Team

Valve rates this game Steam Deck Verified.

Be the first to comment on Vaudeville.