15th Speech in Noise Workshop, 11-12 January 2024, Potsdam, Germany 15th Speech in Noise Workshop, 11-12 January 2024, Potsdam, Germany

P44Session 2 (Friday 12 January 2024, 09:00-11:30)
Exploring the effect of semantic context in dynamic cocktail-party listening

Moritz Wächtler, Hartmut Meister
Jean-Uhrmacher-Institute for Clinical ENT-Research, University of Cologne, Germany
Faculty of Medicine and University Hospital Cologne, Department of Otorhinolaryngology, Head and Neck Surgery, University of Cologne, Cologne, Germany

Cocktail-party situations are common in everyday conversation. They can either be static (target talker remains the same) or dynamic (target changes unpredictably). Due to the need to monitor multiple talkers and to switch attention from one talker to another, dynamic situations are associated with a higher cognitive load and thus a decrease in speech recognition performance relative to static situations, referred to as “costs” (see Lin & Carlile, 2015; Meister et al., 2020). However, the corresponding studies typically used matrix sentences which, due to their nonsensical nature, offer only little semantic context information compared to conversations from everyday life. As there is evidence that semantic context aids stream segregation and word recall (e.g., Meister, 2013), we assume that context effects can alleviate cognitive demand. Against the background that cognitive resources are limited, we hypothesize that higher context allows listeners to allocate more cognitive resources to the challenges of dynamic cocktail-party listening and therefore leads to lower costs.

Cocktail-party situations with three competing talkers located at different positions were simulated. Listeners were asked to repeat back words from the target talker, which had a higher voice than the two masker talkers. The listening situation was dynamic, meaning the target talker changed its position in unpredictable pseudo-random patterns. The three talkers either uttered matrix sentences such as “Simon orders forty wet bottles.” that are assumed to have a low predictability (low context) or more meaningful everyday-life sentences like “The hen laid an egg.” (high context).

The low and high context sentences were based on established speech tests for the German language. However, we did not use the original audio recordings of those tests but synthesized all sentences using a text-to-speech algorithm. In this way, we could improve comparability between high and low context materials by avoiding the talker differences present in the original recordings. In addition, this allowed us to extend the set of words used to form the matrix sentences, thus reducing the listener’s chance of guessing words typically inherent in closed-set material. Preliminary data from young normal-hearing listeners will be shown. The results will be analyzed using mathematical metrics for linguistic context effects and will be discussed against the background of cognitive speech recognition models.

Funding: Deutsche Forschungsgemeinschaft (ME2751/3-2).


Last modified 2024-01-16 10:49:05