Full-duplex conversational audio datasets

License two-person full-duplex conversations recorded at 48 kHz, with each speaker captured on a separate channel at the source. The synchronized tracks preserve the timing of pauses, turn changes, and overlapping speech.

48 kHz · 16 Bit

Why full-duplex data matters

Train models to manage a turn

Voice models that listen while speaking need examples of how people time their responses. Separate tracks let a model learn when to keep talking through a brief “mm-hmm,” wait through a hesitation, or yield to an interruption.

Test how agents respond to overlap

A listener can correct a detail while an agent is still speaking. With separate tracks, you can pinpoint the cue and measure whether the agent incorporates it, keeps going, or hands over the turn.

See the Duplex Cue benchmark

Off-the-shelf conversation datasets

Natural conversation

Unscripted dialogue between two people across topics, accents, and everyday recording environments. Use it to train and evaluate how a model handles natural pacing, hesitations, and interruptions.

Languages
19 languages
Example topics
Childhood games, phones in daily life, personal style, and cooking

Agent–customer roleplay

Two people play an agent and a customer, from the opening question through resolution. Use it to build and test how voice agents handle multi-turn tasks.

Languages
English, Korean, Japanese, Mandarin Chinese, and Cantonese
Example scenarios
Travel and booking, financial services, healthcare scheduling, and technical support

Inventory-wide language coverage

  • Arabic
  • Bengali
  • Cantonese
  • Dutch
  • English
  • Finnish
  • French
  • German
  • Greek
  • Gujarati
  • Hindi
  • Hungarian
  • Indonesian
  • Italian
  • Japanese
  • Korean
  • Malay
  • Mandarin Chinese
  • Marathi
  • Najdi Arabic
  • Polish
  • Portuguese
  • Punjabi
  • Romanian
  • Russian
  • Spanish
  • Swahili
  • Tagalog/Filipino
  • Tamil
  • Telugu
  • Thai
  • Turkish
  • Ukrainian
  • Urdu
  • Vietnamese
  • Yoruba

Inventory-wide accent coverage

  • American (California)
  • American (General)
  • American (Midwest)
  • Arabic (Modern Standard)
  • Argentina
  • Brazil
  • British (Northern)
  • British (RP/Standard)
  • Canada
  • China
  • Colombia
  • Egypt
  • Germany
  • Great Lakes / Inland North
  • India
  • Indonesia
  • Japanese (Tokyo)
  • Kenya
  • Korean (Busan)
  • Korean (Seoul)
  • Malaysia
  • Mexico
  • Mid-Atlantic
  • Midland / Central U.S.
  • New England
  • New York / New Jersey
  • Nigeria
  • Philippines
  • Singapore
  • Southern
  • Texas / Southern Plains
  • Upper Midwest
  • Venezuela
  • Vietnam
  • Western U.S.