DuplexChat

large-scale, two-speaker, full-duplex spoken-dialogue corpus built from public podcast feeds. This repository distributes no audio

Contents

manifest_counts.json holds the same totals.

Row schema

Each line is one two-speaker dialogue clip:

{
  "language": "en-us",
  "rss_url": "https://.../podcast/rss",
  "audio_url": "https://.../episode.mp3",
  "dialogue_idx": 1,
  "episode_start_sec": 80.22,
  "episode_end_sec": 708.44,
  "duration_sec": 628.22,
  "speakers": ["SPEAKER_04", "SPEAKER_02"],
  "episode_duration_sec": 2165.52
}

• rss_url — the source podcast feed.
• audio_url — the source episode audio (download target).
• episode_start_sec / episode_end_sec — the dialogue span within the episode.
• duration_sec — span length.
• speakers — the two diarized speaker labels for the clip.

Rows are deduplicated globally by (audio_url, dialogue_idx).

Reconstruction

1. Download the episode at audio_url.
2. Slice the segment [episode_start_sec, episode_end_sec].
3. Re-run the speech separation/restoration to recover one-speaker-per-channel stereo audio.

The construction pipeline is released separately.

License and copyright

MIT Note that all rights regarding the audio content and RSS content belongs to the respective right holders.