Acoustic parameter combinations underlying mapping of pseudoword sounds to multiple domains of meaning: Representational similarity analyses and machine-learning models.

In spoken language, iconicity, referring to the resemblance between the sound structure of words and their meaning, is often studied using pseudowords. Previously, we showed that representational dissimilarity matrices (RDMs) of the shape ratings of pseudowords correlated significantly with RDMs of acoustic parameters reflecting spectro-temporal variations; the ratings also correlated significantly with voice quality parameters. Here, we examined how perceptual ratings relate to these parameters of pseudowords across eight meaning domains. We largely replicated our previous findings for shape, while observing different patterns for other domains. Using a k-nearest-neighbor (KNN) machine-learning algorithm, we compared 4095 combinations of 12 acoustic parameters (three spectro-temporal and nine characterizing vocal quality) to determine the optimal combination associated with iconicity ratings in each domain. We found that iconic mappings were linked to domain-specific combinations of acoustic parameters. One spectro-temporal parameter, the fast Fourier transform, contributed to all domains, indicating the importance of time-varying spectral properties for iconicity judgments. We applied the KNN approach to generate shape ratings for 160 real words. These generated ratings strongly correlated with perceptual ratings of real words, indicating the value of the KNN approach to assess iconic mapping in natural languages. Our findings support the relevance of iconicity to language.