In the sixth episode of the ELOQUENCE podcast, the focus was on one of the most important challenges in today’s AI landscape: how to make artificial intelligence systems more interpretable, trustworthy, and useful in real-world dialogue scenarios.
The guest was Dr Petr Motlíček, Senior Research Scientist at the Idiap Research Institute in Switzerland and Associate Professor at Brno University of Technology. With a strong background in speech processing, spoken language understanding, and AI-based communication technologies, Petr shared insights into his research journey and his work within the ELOQUENCE project.
Petr’s path into AI began during his studies in electrical engineering at Brno University of Technology. His early research focused on data-driven models for speech communication, particularly speech coding — the process of encoding and transmitting human voice as efficiently as possible. Over time, as the field evolved from pattern recognition and machine learning toward today’s broader understanding of artificial intelligence, his work also expanded toward speech processing, spoken language understanding, and dialogue modelling.
A central theme of the episode was the meaning of interpretable and trustworthy AI. Petr explained that many AI systems are still used as black boxes: users provide input, receive output, but often do not know how the system reached its conclusion, what information it used, or whether the result should be trusted. In sensitive applications, this is not enough. AI systems need to provide explanations, reliability indicators, and, ideally, the ability to recognise when they do not know the answer.
This is especially important in high-risk scenarios. Within ELOQUENCE, one of the key examples is the use of AI in call centre environments, including emergency call analysis. In such cases, incorrect information can have serious consequences. If an AI system extracts the wrong address from an emergency call, for example, this could directly affect the response process. For that reason, human involvement remains essential. AI should support operators and managers, not replace their judgement.
The episode also explored the idea of knowledge infusion, where AI systems are connected with reliable internal databases or human expertise. This can help reduce hallucinations and ensure that generated outputs are grounded in trusted information. In practice, this means designing systems that combine machine intelligence with human oversight and domain-specific knowledge.
Petr also discussed the work carried out in ELOQUENCE during the first year, including the development of resources for dialogue modelling and the creation of models that can support downstream applications, such as assessing whether a call was resolved successfully or whether human intervention may be needed.
Looking ahead, Petr highlighted several important directions for trustworthy AI, including causality, bias mitigation, and AI regulation. He emphasised that future systems should not only provide outputs, but also help users understand why those outputs were produced and whether they can be relied upon.
The conversation offered a clear message: the future of AI in dialogue systems depends not only on performance, but on trust, transparency, and meaningful collaboration between humans and machines.
Listen to the full episode here.
