Research

Research Projects

Selected research projects at the intersection of AI, scientific communication, NLP, and digital media.


Selected Work

2026

Personal Project

Cinematic Speech as a Stress Test for Facial Animation Systems

An NLP and Computer Vision study of emotional bias in speech-driven facial animation benchmarks

Sole researcher

This project investigates a key limitation of speech-driven facial animation: the reliance on phonetically balanced datasets that lack emotional diversity. It introduces cinematic dialogue as a stress test, comparing VOCASET with a curated Disney Cinematic Corpus and extending the analysis to model expressiveness.

Research Focus

Current systems achieve accurate lip synchronization but limited emotional expressiveness, likely due to the neutral bias of training data.

Methods

  • VOCASET vs. cinematic corpus (73 dialogues, 7 films)
  • Emotion classification + human annotation
  • Human–model agreement and distribution gap
  • MediaPipe landmarks for facial expressiveness

Outputs

  • Evidence of strong neutral bias (68.5% vs. 32.9%)
  • 31.5% human–model agreement on cinematic dialogue
  • Multimodal analysis linking NLP and facial animation

2026

Research Internship Project

SciTeller: An LLM-Based Framework for Persona-Adaptive Scientific Storytelling

A two-stage framework for generating audience-adapted scientific narratives

EURECOM · Politecnico di Torino — Andrea Sillano, Pasquale Lisena, Raphaël Troncy, Tommaso Calò, Luigi De Russis

Lead developer — dataset construction, framework design, model fine-tuning, evaluation, and web application

SciTeller is a two-stage framework that transforms scientific papers into audience-adapted narratives by separating content planning (Splitter) from generation (Storyteller).

Research Focus

Existing systems generate generic summaries with limited control over audience adaptation and weak guarantees of faithfulness.

Methods

  • Splitter + Storyteller pipeline
  • 62 papers, 190 stories with persona annotations
  • Section-level semantic alignment
  • Evaluation via StoryScore (grounding, coherence, hallucination control)

Outputs

  • Framework for controllable scientific storytelling
  • Structured dataset and web platform
  • Improved semantic alignment and narrative quality

Next

These projects connect the technical work behind the publications with broader experimentation in AI-generated storytelling, evaluation, and expressive media.