Frontiers in Artificial Intelligence (Jul 2021)
Perception in Black and White: Effects of Intonational Variables and Filtering Conditions on Sociolinguistic Judgments With Implications for ASR
Abstract
This study tests the effects of intonational contours and filtering conditions on listener judgments of ethnicity to arrive at a more comprehensive understanding on how prosody influences these judgments, with implications for austomatic speech recognition systems as well as speech synthesis. In a perceptual experiment, 40 American English listeners heard phrase-long clips which were controlled for pitch accent type and focus marking. Each clip contained either two H* (high) or two L+H* (low high) pitch accents and a L-L% (falling) boundary tone, and had also previously been labelled for broad or narrow focus. Listeners rated clips in two tasks, one with unmodified stimuli and one with stimuli lowpass filtered at 400 Hz, and were asked to judge whether the speaker was “Black” or “White”. In the filtered condition, tokens with the L+H* pitch accent were more likely to be rated as “Black”, with an interaction such that broad focus enhanced this pattern, supporting earlier findings that listeners may perceive African American Language as having more variation in possible pitch accent meanings. In the unfiltered condition, tokens with the L+H* pitch accent were less likely to be rated as Black, with no effect of focus, likely due to the fact that listeners relied more heavily on available segmental information in this condition. These results enhance our understanding of cues listeners rely on in making social judgments about speakers, especially in ethnic identification and linguistic profiling, by highlighting perceptual differences due to listening environment as well as predicted meaning of specific intonational contours. They also contribute to our understanding of the role of how human listeners interpret meaning within a holistic context, which has implications for the construction of computational systems designed to replicate the properties of natural language. In particular, they have important applicability to speech synthesis and speech recognition programs, which are often limited in their capacities due to the fact that they do not make such holistic sociolinguistic considerations of the meanings of input or output speech.
Keywords
