Imagine a voice assistant that doesn't just follow a script but actually understands the nuance of different cultures and languages without being biased. It's like upgrading a basic chatbot to a professional operator who knows exactly when to step back and let a human take over in an emergency. Instead of starting from scratch, it polishes existing AI tools to make them safer and more reliable for everyone in Europe.
Current AI voice assistants often struggle with unstructured conversations, exhibit cultural or gender biases, and lack the reliability required for safety-critical emergency services.
An Interactive Playground for LM validation and a dialogue manager for high-risk use cases. They also created SDialog, a Python library for synthetic dialogue generation.
If you are an emergency call center dealing with high-stress, unstructured dialogues — this project developed a dialogue manager for high-risk use cases that ensures safety-critical responses with humans-in-the-loop.
If you are a device manufacturer dealing with factual inaccuracies in voice assistants — this project developed information retrieval and fact-checking against online knowledge bases to create reliable home-assistants.
If you are a service provider dealing with cultural bias and language barriers across 10 countries — this project developed bias-controlled language models that align with European values and support all EU languages.
Based on available project data, no specific pricing or cost details are provided as the project focuses on improving open-source LLMs.
Yes, the project includes 6 industrial partners and tests the technology through 4 real-world pilots, including customer service and emergency response.
The project follows Open Science principles, aiming to release data, models, and publications in open access.
The system is specifically designed to mitigate gender, cultural, and racial biases to ensure compliance with EU ethical and legal standards.
It uses a lightweight projector to connect speech encoders with LLMs and a Python library called SDialog for synthetic dialogue generation.
The consortium is heavily weighted toward industrial application, with 6 industry partners (38% ratio) including major players like Telefonica and Omilia. With 16 partners across 10 countries, the group balances academic research (5 universities, 4 research centers) with practical SME agility (5 SMEs), suggesting a strong path from lab to market.
Contact Telefonica Innovacion Digital SL in Spain
Contact us to explore licensing for the bias-control modules.
