Synthetic Pipeline Generates Intent-Labeled Full-Duplex Conversational Speech

A new pipeline synthesizes two-channel, intent-labeled conversational speech from relational event lists to train full-duplex dialogue systems. The method uses LLMs to author conversational acts and aligns independently synthesized speech segments to a shared clock to model turn-taking phenomena.