In this ELOQUENCE webinar, titled “Building Trust in Machine Judgement,” the discussion focused on how AI systems can better understand human interaction and provide responses that are more accurate, transparent and useful in real-world contexts.
The session featured Oldřich Plchot and Martin Fajčík from Brno University of Technology, both actively involved in the ELOQUENCE project. Their work focuses on different aspects of speech processing, language technologies and methods for improving the reliability of AI systems.
The first part of the webinar, presented by Oldřich, explored how moving from text-based inputs to speech-based interaction can make human-machine communication more natural. While text can capture the words a person uses, speech carries much more information. Tone, rhythm, stress, emotion, hesitation and other paralinguistic cues can all influence how a message should be understood. When speech is simply transcribed into text, much of this information is lost.
This is especially important in situations where the user’s emotional state or level of urgency matters. For example, in emergency call settings, a person’s stress level may help determine whether a request should be handled by an AI assistant or immediately redirected to a human operator. By working directly with speech inputs, AI systems can potentially respond in ways that are more sensitive to the user’s situation and needs.
Oldřich also explained how ELOQUENCE is exploring architectures that connect speech foundation models with large language models. Instead of treating speech recognition and language understanding as completely separate steps, these systems aim to link the internal representations of speech and text more directly. This can help reduce latency, preserve more information from the original speech signal and support more natural interaction.
In the second part of the webinar, Martin introduced two research directions focused on making AI responses more useful and trustworthy. The first dealt with helping users understand what information is important in a system’s answer. Modern retrieval systems can search through very large collections of documents, but their results may be long and difficult to process. ELOQUENCE researchers are developing methods that can highlight the most relevant parts of retrieved documents, helping users quickly see why a certain answer or result matters.
The second research direction focused on improving the way language models generate responses. Instead of forcing a model to regenerate everything word by word, the idea is to allow it to copy relevant parts from the input when appropriate and generate only what needs to be changed. This approach could reduce errors, make outputs more traceable and help users better understand where information comes from.
The discussion also addressed real-world adoption. Many of the techniques presented are not purely theoretical; they are already being developed as working systems. However, challenges remain, particularly around multilingual speech data, accents, low-resource languages and explaining AI decisions in ways that users can understand.
The key message of the webinar was clear: building trust in AI is not only about making systems more powerful. It is about making them more understandable, more context-aware and better aligned with the way people actually communicate.
Watch the full webinar here.
