The TOSCA-MP project aimed to develop user-centric content annotation and search tools for professionals in networked media production and archiving (television, radio, online), addressing their specific use cases and workflow requirements. The project brought together 10 partners from 6 European countries including industry partners providing solutions for the media industry, public service broadcasters as well as their European association, a university and research centres. TOSCA-MP investigated scalable and distributed content processing methods performing advanced multimodal information extraction and semantic enrichment. Other key technology areas included search methods across heterogeneous networked content repositories and novel user interfaces. An open standards based service oriented framework integrated the components of the system.
🌍 @lina_conti and @luisabentivogli are heading to #LREC2026 in Palma! They'll present two papers:
📄 "Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation"
Paper link:
🤔 What Matters in Data for DPO? I asked myself this question a few days ago while trying to understand how to generate a dataset with preferences to run #DPO. This recent #NeurIPS paper answered some of my questions. The findings are simple but crucial for data creation:
🎓 Come and join our group! 🎓
We offer 2 fully funded PhD positions:
🌍 Human-Centred Evaluation Frameworks for Multilingual Technologies (A6)
🤖 Multimedia Personalization with Multimodal Large Language Models (A7)
⏰ Deadline: 15 May 2026
🔗 Details: https://iecs.unitn.it/education/admission/call-for-application
Our pick of the week by
@FBKZhihangXie
: "Detecting Hallucination in SpeechLLMs at Inference Time Using Attention Maps" by @JWaldendorf, Bashar Awwad Shiekh Hasan and Evgenii Tsymbalov
📰
#SpeechLLM #Hallucination
🚀 New paper: Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps
📄 http://arxiv.org/abs/2604.19565
🧩 Lightweight inference-time detection for SpeechLLM hallucinations via audio attention.
✨ Attention classifiers beat uncertainty baselines on ASR and S2TT.