Over the past two years, ELOQUENCE has grown from a collection of ambitious research ideas into a tightly connected ecosystem of multilingual, multimodal conversational AI technologies. This progress has been shaped by close collaboration between partners bringing diverse expertise across speech, language, evaluation and real-world deployment.
In this interview, Javier García Gilabert and Miguel Claramunt from the Barcelona Supercomputing Center reflect on their journey within ELOQUENCE so far, from building core technological infrastructure to enabling hands-on experimentation across the consortium, and share their perspective on what has made this collaboration both impactful and inspiring.
Q: Looking back at the first two years of ELOQUENCE, which achievement from your team are you most proud of?
BSC: We developed a suite of interactive “playground” demos covering Salamandra and the Whisper-based pipelines that quickly became a common experimentation hub across ELOQUENCE. These interfaces have supported cross-partner collaboration, accelerated model debugging, and enabled early detection of recurrent issues such as hallucinations, gender biases, and domain drift.
Q: What has been the most rewarding or inspiring part of collaborating with the ELOQUENCE consortium?
BSC: The most rewarding aspect has been working in a genuinely interdisciplinary environment where each partner brings a different perspective on speech and language technologies. Collaborating with teams focused on ASR, dialogue systems, evaluation, and user-centric applications has pushed us to think beyond our own models and design solutions that are robust in real, multimodal workflows.
Q: Which result, insight, or technological advancement do you feel has had the strongest impact on the project so far?
The most impactful result so far has definitely been the Interactive Playground. It’s the first time all ELOQUENCE models, tools, and modalities come together in one place, and it has completely changed how partners work with the technology. Instead of each group testing things in isolation, the Playground gives everyone a shared space to try models, spot issues quickly, compare behaviours, and plug in pilot-specific data. It has basically become the project’s “main core,” making it much easier to debug, iterate, and turn research ideas into something that actually works across the pilots.
Q: What is one thing that worked particularly well in your work package or activity and why?
One thing that worked particularly well was the close loop between WP2’s model development and WP5’s Interactive Playground. The Playground gave WP2 an immediate way to test new features like speech connectors, RAG pipelines in realistic, pilot-like conditions.
Q: Which moment, milestone or breakthrough from the past two years stands out as especially meaningful for your organisation?
For us, the milestone that really stands out is the first full deployment of the Interactive Playground with live ELOQUENCE models running end-to-end. It was the moment when all the scattered pieces (speech encoders, LLMs, RAG components, pilot data, feedback mechanisms) finally came together in a single working system. Seeing our models respond in real time, grounded with pilot-specific knowledge and accessible to all partners, made the project feel “real.”
Q: What strengths or expertise did your team bring that you believe contributed most to the project’s success?
We contributed most through a combination of technical depth and fast, hands-on integration work. Our team brought strong expertise in multilingual LLMs, and development integration in high-performance environments, which allowed us to quickly adapt models, debug complex behaviours, and turn research outputs into usable components for the pilots.
Q: What have you learned through ELOQUENCE that will be valuable for your future work or research?
One of the biggest lessons has been how essential tight integration and early real-world testing are when developing conversational AI. Working across partners, languages, modalities, and use cases made it clear that offline benchmarks only tell part of the story, systems reveal completely different behaviours once they interact with real users, domain knowledge, and messy dialogue histories.
Q: What aspect of the project’s final year are you most excited about and why?
We are most excited about the final year because it’s when the project becomes truly interactive: with SDialog giving us a way to generate realistic, persona-rich conversations and the feedback pipeline finally in action, we’ll be able to test models in real-world-like scenarios and actually use user feedback to refine them on the fly. Together, these pieces turn the Playground into a genuine learning loop, where systems improve based on how people engage with them, making this last phase feel both dynamic and hugely impactful.
Q: Which project result or development do you believe has the greatest potential for long-term impact beyond ELOQUENCE?
We think the result with the strongest long-term impact is the Interactive Playground as a unified experimentation and feedback ecosystem. It isn’t just a demo platform – it’s a reusable framework that brings together multilingual LLMs, speech models, RAG pipelines, persona simulation (via SDialog), and structured feedback collection in one place.
Q: From your perspective, what do you see as the most important opportunities and responsible AI challenges emerging right now? How do you think ELOQUENCE can help address them?
Right now there’s a huge opportunity to build conversational systems that actually understand speech, context, and different languages well enough to be useful in real services–that also comes with great responsibilities: minimizing hallucinations and biased behaviours; keeping models grounded and predictable. What’s exciting about ELOQUENCE is that it directly tackles these issues: with multilingual models, bias-focused evaluation, RAG grounding, SDialog for stress-testing, and a full feedback loop to catch problems early, it gives us a practical way to build more trustworthy and reliable systems from the start.
Q: If you had to describe ELOQUENCE’s journey so far in one sentence, what would it be?
ELOQUENCE has been a fast, collaborative push from scattered research ideas to a fully integrated ecosystem where multilingual, multimodal models can actually be tested, grounded, and improved in realistic conditions.
Looking ahead, BSC’s vision underscores what ELOQUENCE ultimately aims to achieve: conversational AI systems that are not only powerful and multilingual, but also grounded, transparent, and shaped by real-world use – turning research innovation into technology that can be trusted and applied beyond the project itself.
