The effect of perceived human-likeness on voice-user interface–directed speech

Abstract

Many interact daily with voice–user interfaces (VUIs), but acoustic research on VUI-directed speech (VDS) is relatively new. Prior work indicates intensity and F0 correlate with VDS [Cohn et al., 2022, JPhon 90]. Multiple acoustic variables of VDS were analyzed to explore if VDS is a register distinct from human-directed speech (HDS) and whether perceived human-likeness of VUI voices affects VDS characteristics. 27 participants’ Zoom recordings of 13 pre-scripted prompts and pre-recorded responses from two Amazon AWS-Polly-generated VUI voices (rated for human-likeness independently and by participants) were acoustically analyzed for word-initial voiceless plosives voice onset time (VOT), pitch variation, and vowel quality and quantity. Results of linear mixed-effects models indicate evidence of VDS-specific acoustic characteristics, some of which are affected by participants’ perceived human-likeness of the voices. Differences in pre-exposure and VUI interactions occur for /p/ VOT in consonant clusters, vowel duration (except /ɪ/), and /ɑ/ F2. Statistical differences are found for /p/ VOT in consonant clusters and vowel quality (e.g., /ɑ/ F1 and F2, and /ɪ/ F2) based on perceived human-likeness by participants. This study contributes to the growing VDS work examining how humans speak with devices and what affects VDS, which may influence considerations of VUI voice development.