AI-Driven Podcast Generation from Technical Documents Using Retrieval-Augmented Generation and Multi-Speaker TTS
Authors/Creators
Description
This paper introduces a comprehensive system for automated podcast generation from academic and technical documents using Retrieval-Augmented Generation (RAG) and multi-speaker text-to-speech (TTS) synthesis. The system bridges a critical accessibility gap by combining advanced document processing, vector-based knowledge retrieval, conversational AI, and neural audio synthesis to convert static PDF documents into engaging and accessible podcast content.
The architecture integrates several state-of-the-art components: the Docling framework for advanced PDF processing with multimodal OCR (including table and image extraction), PostgreSQL with pgvector for scalable vector storage capable of handling millions of document chunks, LangChain for robust RAG implementation with semantic chunking, and Gemini 2.5 Flash TTS for high-quality, multi-speaker audio with distinct speaker identities.
A novel conversation generation pipeline creates natural dialogue between two AI personas (Sarah and Mike), incorporating anti-repetition mechanisms, dynamic topic management, and conversational quality control to sustain engagement throughout extended discussions. The system addresses key technical challenges such as RAG scalability via optimized vector indexing and hybrid retrieval, conversational naturalness through contextual awareness and prosodic coherence, and enterprise-level deployment with features like user isolation, real-time processing, and cloud-native architecture using MinIO for distributed file storage.
Our engineering approach prioritizes modularity, reproducibility, and practical deployment, with built-in monitoring and quality assurance. Evaluations show the system effectively processes complex technical documents with multimodal content, generates coherent and engaging AI-driven conversations, and produces broadcast-quality podcast audio—while maintaining scalability and cost-efficiency for enterprise use.
This work marks a significant step forward in AI-driven content accessibility, offering a scalable solution to transform technical documentation into audio content for broad audiences.