Shobhit Banga and the VoiceArena team have launched a new data collection called Joy, designed to improve how artificial intelligence handles the timing and flow of natural human speech. The dual-channel dataset contains 100,000 hours of unscripted conversations featuring 50,000 speakers across 10 languages, recording each participant on a separate track to preserve the authentic structure of live dialogue.
The release targets a limitation in current voice AI models that are typically trained on synthetic dialogue or clean scripts, which omit the interruptions and verbal fillers common in real interaction. By marking overlaps and prompts in the data, VoiceArena aims to teach models the nuances of conversational timing, stating that this first release is only the beginning of their data collection efforts.
