At this year’s Interspeech 2025, the ELOQUENCE consortium once again demonstrated the strength of European collaboration in the field of speech and language technologies. Partners from Foundation Bruno Kessler (FBK) and the Brno University of Technology (BUT) brought forward several scientific contributions that reflect the project’s central goals – developing speech technologies that are multilingual, trustworthy and inclusive.
Exploring New Frontiers with FBK
The FBK team, including Seraphina Fong, Marco Matassoni and Alessio Brutti, participated in the conference with two papers that captured the attention of the research community.
One of them, “Scaling and Enhancing LLM-based AVSR: A Sparse Mixture of Projectors Approach,” introduced an innovative way to enhance audio-visual speech recognition by leveraging the capabilities of large language models. The approach aims to make multimodal systems more efficient and adaptable, an important step towards human-like interaction between users and machines.
The second paper, “Speech LLMs in Low-Resource Scenarios: Data Volume Requirements and the Impact of Pretraining on High-Resource Languages,” presented as a poster by PhD student Seraphina Fong, examined how speech models trained on high-resource languages perform when adapted to low-resource settings, a key topic for achieving fairness and inclusivity in AI.
Pushing the Boundaries of Conversational AI with BUT
Researchers from Brno University of Technology (BUT) also made a strong impact, presenting papers that address different aspects of speech understanding and dialogue management.
Their oral presentation on “Factors affecting the in-context learning abilities of LLMs for dialog state tracking” explored how large language models can dynamically adapt to conversational context, an essential feature for next-generation AI assistants. Another study, “Approaching Dialogue State Tracking via Aligning Speech Encoders and LLMs” , presented as a poster, focused on improving the way AI systems maintain awareness of dialogue flow, making human-computer interactions more coherent and natural.
MLC-SLM Challenge Workshop
The ELOQUENCE presence extended beyond the main conference sessions – both FBK and BUT contributed to the MLC-SLM Challenge Workshop, with contributions to two different tracks of the challenge. Each team presented its own system developed within ELOQUENCE. Marco Matassoni (FBK) showcased a system co-developed with Telefónica and the University of Essex, while the BUT team presented their “BUT System for the MLC-SLM Challenge”, focusing on multilingual and multi-speaker speech recognition.
A Shared Vision for Speech Technology
The work of FBK and BUT exemplifies the depth of knowledge and shared ambition within ELOQUENCE. Through a combination of cutting-edge LLM research and multilingual, low-resource innovation, the project is paving the way toward speech technologies that are not only powerful, but also fair, inclusive and truly global.
For all publications and more in-depth information, visit our Publications page.
