Karamel’s Adventures: Building an AI-Powered Multilingual Storybook Generation Pipeline

This paper presents a fully automated pipeline for converting monolingual, illustrated PDF storybooks into multilingual, AI-narrated interactive digital publications. The system was developed to disseminate 53 children's storybooks—originally produced in English by the Houston Education Attaché Office of the Republic of Türkiye and hosted at storiesofturkiye.com—across 34 target languages, covering the cultural, historical, and geographical heritage of Türkiye for young readers worldwide. The pipeline comprises four sequential stages: (1) structured PDF decomposition into text and image assets using PyMuPDF, (2) context-aware translation and editorial refinement via a locally hosted large language model (LLM) running under LM Studio, (3) multilingual text-to-speech (TTS) synthesis with optional zero-shot voice cloning using the Chatterbox model, and (4) automated generation of flip-book–style HTML5 web publications. The resulting system produces 15 languages with full audio-text output and an additional 19 languages with text-only output, reaching over 34 distinct linguistic communities through the diplomatic education network of Türkiye's overseas representations. We describe the architectural decisions, prompt engineering strategies, AI hallucination mitigation, and cross-lingual voice transfer challenges encountered, and we reflect on the broader implications of LLM-driven educational content localisation at scale.