JUMAS addresses the need to build an infrastructure able to optimise the information workflow in order to facilitate later analysis. New models and techniques for representing and automatically extracting the embedded semantics derived from multiple data sources will be developed. The most important goal of the JUMAS system is to collect, enrich and share multimedia documents annotated with embedded semantic minimising manual transcription activity. JUMAS is tailored at managing situations in which multiple cameras and audio sources are used to record assemblies in which people debates and event sequences need to be semantically reconstructed for future consultations. The prototype of JUMAS will be tested interworking with legacy systems, but the system can be viewed as able to support business processes and problem-solving in a variety of domains.
3️⃣ "Cross-Attention is Half Explanation in Speech-to-Text Models"
👥 @sarapapi, @DennisFucci, @mgaido91, @negri_teo, @luisabentivogli
🇪🇺 DVPS EU project
📄
1️⃣ "Do What I Say: A Spoken Prompt Dataset for Instruction-Following"
👥 @MaikeZufle, @sarapapi, Fabian Retkowski, Szymon Mazurek, @mkasztelnik, Alexander Waibel, @luisabentivogli, @_janius_
🇪🇺 Meetween EU project
📄
Our pick of the week by
@lina_conti
: "Greater accessibility can amplify discrimination in generative AI" by
@CarolinHolterm, @minhducbui_nlp, @KaitlynZhou, @vjhofmann, @kelina1124, @anne_lauscher
📰
#GenderBias #SpeechLLM
Pick of the week @fbk_mt: "Greater accessibility can amplify discrimination in generative AI"
Gender bias in speech-based LLMs examined from multiple angles: a user survey, automatic bias measurement, and pitch manipulation experiments.
https://arxiv.org/pdf/2603.22260