Introduction
The creation of professional presentations is a multi-faceted task involving content synthesis, visual design, and narrative rehearsal. Despite the proliferation of AI tools, existing "document-to-slide" solutions often struggle with domain specificity, content redundancy, and the "delivery gap"—the lack of integrated oral narration. To address these challenges, we present PASS (Presentation Automation for Slide Generation and Speech), an end-to-end pipeline designed to transform raw documents into complete, narrated presentations.
The PASS Architecture
The PASS framework is divided into two core modules: Slide Generation and Slide Presentation. This separation ensures that the visual content is logically structured before the narrative layer is added.
Slide Generation
The generation process begins with the extraction of text and images from a source document. An LLM then generates 8–10 logical slide titles tailored to the target audience's technical level. Following title generation, the system extracts relevant snippets and summarizes them into concise bullet points. A key innovation in PASS is the Image Mapping step, which uses multimodal analysis to automatically place figures from the source document onto the most contextually relevant slides.
Slide Presentation
Unlike traditional tools, PASS extends into the delivery phase. A Presenter Script Generator creates a natural-sounding narrative for each slide based on its bullet points. This script is then processed by a Tacotron-2 based Text-to-Speech (TTS) engine to produce high-quality audio voiceovers, which are synchronized with the visual slides to create a seamless presentation experience.
Evaluation and Results
The PASS pipeline was evaluated using the SciDuet dataset, comprising 81 research papers from ICML and NeurIPS. To move beyond traditional metrics like ROUGE, the authors employed G-Eval with Llama-3-70B to assess coherence, redundancy, and relevance.
Results indicate that PASS significantly outperforms existing state-of-the-art models. Specifically, GPT-PASS achieved a coherence score of 8.79/10, compared to 7.42 for the D2S baseline. Notably, the system excelled at reducing redundancy, scoring 8.35/10 while older models lagged at 6.11. Overall, the integrated pipeline achieved an average performance score of 8.96/10, demonstrating its effectiveness in creating high-quality, non-repetitive presentations.
Future Directions
While PASS represents a significant advancement in presentation automation, future work will focus on enhancing interactivity. Integrating Retrieval-Augmented Generation (RAG) could allow the AI presenter to handle real-time Q&A sessions. Additionally, providing users with more control over the "persona" of the AI voice and conducting large-scale human studies will be crucial for further refining the engagement and professional utility of the system.
The code and data for PASS are available at https://github.com/AggarwalTushar/PASS.
Summary updated