On the relation between lexical and prosodic cues of emotion in English: a preliminary corpus-based analysis

Abstract

Abundant literature has shown that emotional speech is characterized by various cues at different linguistic levels. This contribution investigates the relation between lexical and prosodic cues of emotions in English using audiobook recordings from the LibriSpeech corpus. In order to quantitatively evaluate lexical cues of emotions, we computed sentiment analysis scores from audiobook texts using tools and lexicons provided by the Python-NLTK library (i.e. Vader and SentiWordNet), at word and sentence levels. In order to analyse prosodic cues of emotions with a quantitative approach, we extracted several acoustic parameters indicating variations in pitch, voice quality and rhythm. We explored the relation between lexical and prosodic metrics via linear mixed-effects models and found small but significant effects. In particular, pitch parameters seem to be related to the valence of sentences and words, while pitch, rhythm and voice quality seem to be related with the degree of positivity/negativity of sentences and words. However, our models explain but a small amount of variance in the data. We therefore argue that the relation between lexical and prosodic cues is not extremely tight in audiobook recordings, and we speculate that this claim may extend to cases in which the speaker conceptualises/describes emotions. A tighter relation between lexical and prosodic cues may be reserved for cases in which the speaker expresses emotions and/or feels more directly involved. At any rate, our result should be considered as preliminary, given the limitations of an entirely automatic analysis.

Keywords

• voice quality
• Prosody
• emotional speech
• sentiment analysis