Hybrid Approaches for Conversation Group Identification: Integrating Wireless, Audio, and Contextual Information

This paper proposes a novel method for extracting conversation groups by fusing wireless and audio data. We first use wireless signals to identify groups of people that walk together. Then we propose a novel approach for detecting conversation groups using audio data. Our approach examines non-contextual and contextual analyses to align with user permission levels concerning data utilization. We also incorporate turn-taking likelihood, frequent words, topic similarity, and Named Entity Recognition (NER) to identify conversation groups effectively. To further mitigate privacy issues, we propose to use vector feature representations of speech patterns, a subset of relevant keywords, and a vector space representation of the discussed topic. We first convert the audio files into text and analyze the content by identifying frequent words and topics and extracting NER information on the user's phone. Only the featured vectors are received at the third-party server to compute different similarities among conversation participants. By integrating these similarity values, we obtain the non-contextual and contextual-level Similarity between two users, which allows us to determine if they belong to the same conversation group. Experimental results demonstrate that turn-taking likelihood can identify groups by 90%. However, contextual information can achieve high accuracy rates of up to 96.57% and an average F1-score of 94.77%. These findings indicate that our system is highly effective in analyzing data in audio streams for conversation grouping, with potential applications in various social scenarios.