CoLoSS: Cognitive Load Corpus with Speech and Performance Data from a Symbol-Digit Dual-Task

CoLoSS: Cognitive Load Corpus with Speech and Performance Data from a Symbol-Digit Dual-Task Robert Herms1, Maria Wirzberger2, Maximilian Eibl1, G¨unter Daniel Rey2 1Media Informatics, Chemnitz University of Technology, Germany 2Psychology of Learning with Digital Media, Chemnitz University of Technology, Germany {robert.herms, maximilian.eibl}@cs.tu-chemnitz.de Abstract In this paper, a new corpus named CoLoSS (Cognitive Load by Speech and performance data in a Symbol-digit dual-task) is presented, which contains speech under cognitive load recorded in a learning task scenario. In order to obtain a reference for cognitive load, a dual-task approach was applied, including a visual-motor primary task that required subjects to learn abstract symbol combinations and an auditory-verbal secondary task to measure the load imposed by the primary task. We report the methodology of collecting the speech recordings, constructing the corpus and describe the properties of the data. Finally, effects of cognitive load on prosodic as well as voice quality features are investigated in conjunction with the corpus. In its current version, the corpus is available to the scientific community, e.g., for exploring the influence of cognitive load on speech or conducting experiments for speech-based cognitive load recognition. Keywords: speech corpus, speech features, cognitive load, learning performance, dual-task 1. Introduction The human cognitive system is characterized by capacity limitations in information processing (Plass et al., 2010). They refer to human working memory, which provides tem- porary storage and manipulation of information (Baddeley, 1992). Cognitive load is generally considered as the load imposed on an individual’s working memory by a particular (learning) task (Paas and Van Merri¨enboer, 1994). Accord- ing to the cognitive load theory (Sweller et al., 2011), the degree of cognitive load influences the amount and com- plexity of learned content (Paas et al., 2003). Speech databases that include audio recordings of speak- ers under varying levels of cognitive load are rather rare and often created for own research purposes. Different task designs that were developed to investigate the limitations of human working memory in conjunction with speech param- eters can be found in the literature; for instance, reading- comprehension (Yin et al., 2007), Stroop interference (Yap et al., 2010), arithmetic abilities (Gorovoy et al., 2010), and driving under cognitive load (Boril et al., 2010). Moreover, the Cognitive load with Speech and EGG (CLSE) database was created by (Yap, 2012), which includes speech record- ings of subjects participating in three different tasks: Stroop test with time pressure, Stroop test with dual-task, and reading span task. According to the INTERSPEECH 2014 Computational Paralinguistics Challenge (Schuller et al., 2014), the partitioned form of the CLSE database is avail- able for research purposes. Nevertheless, there is still no speech-based corpus available for either the consideration of cognitive load in a learning context or a more sensitive approach to the traditional classification problem. In this paper, CoLoSS (Cognitive Load by Speech and per- formance data in a Symbol-digit dual-task) is introduced— a new corpus that includes speech under cognitive load recorded in a learning task scenario. Compared to existing works in the literature concerning speech-based …

Content truncated to 500 words