From intended topics to emergent discourse: a duration-weighted framework for quantifying conversational drift in X Spaces

Abstract Live social audio platforms such as X Spaces enable conversations to evolve beyond their creator-defined thematic intent, challenging the reliability of static metadata labels for representing real-time discourse. This study investigates conversational drift and thematic stability through a duration-weighted framework combining semantic segmentation, topic classification, and thematic aggregation. Analysing 335 X Spaces sessions comprising approximately 566 hours of conversational audio, we quantify the alignment between intended topics and emergent discussion content based on aggregate thematic duration. The findings suggest that conversational drift follows structured, topic-dependent patterns rather than random disruption. Although off-topic discussion occupied a substantial proportion of conversational time, drift frequently remained semantically adjacent to the intended scope, indicating gradual thematic expansion rather than complete topical deviation. Several recurrent thematic transition patterns were observed between niche technical discussions and broader interpersonal categories. These findings highlight the limitations of static creator-assigned metadata for representing evolving live discourse within synchronous social audio environments. More broadly, they demonstrate the importance of duration-aware approaches for content discovery, recommendation, and moderation in environments where dominant conversational themes may diverge from their originally intended scope.