Edinburgh Napier University

  Audio messaging and voice-based interactions are growing in popularity. Lexical features of a manually-curated dataset of real-world audio tweets, as well as text and video/image tweets from the same user accounts, are analyzed to explore how user-generated audio differs from text. The toxicity, sentiment, topic and length of audio tweet transcripts are compared with their accompanying text, date-matched text tweets from the same users and date-matched video/image tweets and their accompanying text. Audio tweets were significantly less toxic than both text tweets and text that accompanied the audio tweet, as well as significantly lower sentiment than their accompanying text. The topics and word counts of audio, text and video/image tweets also differed. These findings are then used to derive design implications for audio and conversational agent interaction. This research contributes preliminary insights about audio social media messages that may help researchers and designers of audio- and agent-based interaction better understand and design for different media formats.