English (U.S.) conversations in-studio speech dataset

Authentic, unscripted conversations

Unscripted, free-form conversations between two speakers on given topics and role-play scenarios (e.g., between airline customer support and a customer about flight delay and compensation) that captures natural linguistic features such as false starts, filler words and spontaneous turn-taking.

Rich speaker and scenario diversity

Covering a broad range of conversation types and topics, the dataset captures rich variation in language use, tone, pacing and speaker interaction, helping models learn to generalize across different conversational contexts, speaker personalities and interaction styles.

High-fidelity audio with acoustic control

Recorded in rooms at least 8' x 10' x 7' in size, with reverberation time (RT60) under 0.4 seconds at all frequencies, for minimal echo and background noise. Audio is delivered in 48 kHz, 24-bit mono .wav format, with 1-second silence padding and utterances normalized to -26 dB RMS active speech.