About — ELOQUENCE

About ELOQUENCE

AI that speaks
every language.
Earns every trust.

We build multilingual, bias-controlled conversational AI for the environments where getting it wrong isn't an option — healthcare, enterprise, smart home, and beyond.

35+

EU languages
supported natively

Industries with
live deployments

+50%

Customer satisfaction
target, enterprise pilot

50%

Bias reduction
target across pilots

What we're building

Conversational AI that
works for all of Europe.

Most AI is built for English, tested in low-stakes chat, and deployed without accountability. ELOQUENCE takes the harder path — multilingual by design, safety-critical from day one, bias-controlled by architecture, and EU AI Act compliant before it's required.

Multilingual by design 35+ EU languages natively trained — not translated. Including low-resource languages underserved by mainstream AI.

Built for safety-critical environments Healthcare triage, emergency call centres, enterprise customer service — where accuracy and human oversight are non-negotiable.

Active bias mitigation Not just bias-aware — bias-controlled. Measurable reduction targets across gender, culture, ethnicity, and religion.

Privacy-first architecture Federated learning, local processing, zero raw data sharing. Compliance built in — not bolted on.

Multilingual LLMs

EuroLLM · Salamandra · Qwen2-Audio

Foundation models covering all 24 official EU languages — plus regional and low-resource ones.

Retrieval-Augmented Generation

Grounded in your documents

Answers pulled from verified enterprise sources. No hallucination, no guesswork.

Voice & Audio

ASR · Speaker diarization · Speech-to-text

Full audio pipeline for call centres and voice-first environments — not just text chat.

Bias Detection

Synthetic personas + multilingual eval

Surfaces hidden bias before it reaches users. Benchmarked, measurable, auditable.

Federated Learning

Model updates only — no raw data leaves the device

Purpose-built for smart home and consumer environments where privacy is paramount.

Use cases

Four industries.
One platform.

Validated in four real-world environments — each chosen because it demands the highest standards of accuracy, language coverage, and ethical accountability.

🏥 Healthcare

Emergency Call Centre AI

LLM triage layer for healthcare call centres — collects structured clinical info and hands off context summaries to medical staff before the call connects.

150 simulated dialogues

🏢 Enterprise

Self-Adapting Customer Service

RAG-powered agents for finance, insurance, and utilities — generalise to new enterprise contexts without fine-tuning. Empathetic, factually grounded, multi-turn.

+20% first-call resolution

🏠 Smart Home

Privacy-Preserving Voice Assistant

Federated learning for smart home platforms — including Telefónica AURA. Models improve collaboratively without any raw data leaving the device.

Zero raw data transfer

⚖️ Advisory AI

Bias-Controlled Career Guidance

Advisory agent for career counselling that detects and mitigates cultural, gender, ethnic, and religious bias — evaluated across 24+ EU languages.

50% bias reduction target

Impact & outcomes

Numbers that matter.

Measurable outcomes from live pilots and validated project targets — not projections.

+20%

First-call resolution rate

Enterprise pilot target (KR11)

+50%

Customer satisfaction improvement

Enterprise pilot target (KR13)

+20%

Call containment rate

Enterprise pilot target (KR14)

50%

Bias reduction across all pilots

Project-wide target

Healthcare

150+

Simulated clinical dialogues built for the healthcare pilot — including audio, transcriptions, emotional context annotations, and medical ontologies.

Source: D5.2 Pilot 4

Language coverage

35+

EU languages supported — including low-resource languages where mainstream AI falls short. Salamandra covers 35, EuroLLM all 24 official EU languages.

Source: Salamandra-7B, EuroLLM-9B

Open science

Dec '26

Full open release: models, evaluation datasets, multilingual benchmarks, and bias detection toolkits — available to the European AI research and industry community.

Planned: December 2026

Project timeline

36 months. Real milestones.

January 2024 to December 2026 — where we've been and what's coming.

Jan 2024

Project launch

Consortium kickoff in Barcelona. 16 partners, 9 work packages, one goal.

Mid 2024

Core models integrated

EuroLLM, Salamandra, and Qwen2-Audio online. Multilingual RAG architecture finalised.

Late 2024

Interactive Playground live

Public demo at eloquence.lt.bsc.es — test 35+ EU languages, RAG, and audio in real time.

Mid 2025

Midpoint review

All 16 partners reconvened to validate pilot progress and accelerate second-half delivery.

2025–2026

Live pilot deployments

All 4 pilots running in real-world environments — generating measurable, auditable outcomes.

Dec 2026

Open release

Models, datasets, and benchmarks released. Results published. Technology available for adoption.

Leadership

In their own words.

The researchers leading ELOQUENCE — on why this work matters.

Our aim is to develop AI systems that seamlessly integrate into safety-critical applications — enhancing response times and accuracy across Europe's diverse linguistic landscape. This approach ensures AI complements rather than replaces human intelligence.

Dr. Jordi Luque Project Coordinator · Senior Research Scientist, Telefónica
Associate Professor, Universitat Politècnica de Catalunya in LinkedIn

We are leveraging multilingual and multimodal speech foundation models to boost conversational AI across all EU languages — particularly those that are less resourced and have been left behind by mainstream AI development.

Dr. Alessio Brutti Scientific Coordinator · Head of Speech Technology Lab, FBK
Centre for Augmented Intelligence in LinkedIn

The consortium