Submission by CheungSK

Hearsay is a voice-bluffing card game where both sides have to listen for the lie. You face off against an AI interrogator; every move is a spoken claim, and every round you have to decide "is it bluffing?" right back. The AI listens for the acoustic tells in your voice — pauses, fillers, speech rate — but when the AI speaks its own claim, its voice settings shift based on whether it's telling the truth: one persona's voice tightens when lying, another's loosens, another inverts entirely. Three strikes and the Judge reads the verdict. I built four custom AI opponents, each with a distinct voice and a distinct lie-tell signature that YOU have to learn to hear. The Prosecutor is soft-spoken and over-articulated — his tell is almost invisible. The Defendant drops fillers when nervous. The Attorney's voice actually inverts — halting when honest, smooth when lying — so players can't just pattern-match confidence. A hybrid AI brain (deterministic math + Gemini Flash LLM + 2-second deterministic fallback) decides whether to accept or challenge your claim, while a real-time lie-score heuristic derived from Scribe STT metadata gives each persona a different threshold for calling bluff. On the voice side, I used ElevenLabs' Voice Design API to design five custom personas (four opponents + a Clerk tutorial narrator), Scribe Speech-to-Text to extract voice metadata (latency, fillers, WPM, pauses) that drives the lie-detection heuristic on the player, Flash v2.5 TTS with voiceSettings (stability, style, speed) modulated per truth-state so the AI's voice literally sounds different when it's lying, the Music API to pre-generate three tension beds (calm / tense / critical) cached in Vercel KV, and the Sound Effects API for gavel strikes, elimination stingers, and per-persona final-words clips. On AWS Kiro, I built the entire project via spec-driven development: 9 specs covering game-engine, deck-and-claims, voice-tell-taxonomy, ai-opponent, strikes-penalty-system, joker-system, ai-personas, probe-phase, and tension-music-system — each with a requirements doc, design doc, and task breakdown. I shipped five steering files, on-save test hooks, and on-commit changelog hooks — all committed to .kiro/ so the full methodology is auditable in git history. The game-debug MCP is also packaged as a distributable Kiro Power.