Kataribe Studio: AI audio-drama on AMD GPUs

Kataribe Studio is an agentic production pipeline for serialized audio drama, built for the AMD Developer Hackathon ACT III (Track: New Kind of Experience). Audio fiction is booming, but producing even one episode traditionally needs a writers' room, actors, studio time and a mix engineer. Kataribe compresses draft → critique → voices → master into minutes, while keeping the human where taste actually lives: which lines survive, how they sound, what stays canon. How it works: you bring a one-line premise. A scriptwriter agent — running on an open model served from AMD infrastructure — builds an original story bible: cast, speech styles, and a voice profile for every character. One click plans a full episode: scenes, beats, cliffhanger. The scene arrives as structured, performable lines (who / text / emotion / sfx), editable in place. A director agent then scores the scene, cites exact line ids and prescribes concrete fixes; "apply direction" sends those notes back for surgical rewrites that preserve ids. The voice booth casts every character a distinct neural voice, caches lines by content hash, and ffmpeg renders a loudness-normalised master with a room-tone bed. A continuity agent keeps the serial's canon between episodes. AMD integration: every language step — bible, outline, scene, critique, rewrite, continuity summary — is an OpenAI-compatible call to the AMD Token Factory endpoint (AMD Developer Cloud / ROCm). The studio UI shows a live inference meter (calls, tokens, latency) for the episode being produced, so judges can watch AMD inference driving the pipeline. Voice synthesis and mixing run locally so GPU cycles go to language, judgement and creativity. Shipped: hosted demo (kataribe.hikamers.app), public MIT repository with a README that runs, 23 passing tests, a Dockerfile, and a fully voiced episode produced end-to-end by the app itself.