Researchers introduce DuplexChat, an open-source corpus designed to address the lack of speaker-separated data needed for training full-duplex spoken dialogue language models. To create this resource, they developed DuplexChat-Pipe, a pipeline that processes public podcast feeds to extract and separate audio streams.
- The pipeline filters language-specific podcasts, cleans episode audio, and uses diarization-guided extraction to isolate two-speaker dialogue clips.
- Speech separation and restoration techniques are applied to produce distinct channels for each speaker.
- The resulting corpus contains 282,634 hours of English and 132,723 hours of Japanese full-duplex dialogue speech.
- Analysis confirms the dataset exhibits turn-taking dynamics characteristic of human conversations.
This work provides a large-scale, high-quality dataset that enables the training of more effective spoken dialogue models by offering the necessary speaker-separated audio data.