---
license: cc-by-nc-4.0
language:
- en
task_categories:
- automatic-speech-recognition
- text-to-speech
- audio-classification
pretty_name: American English Full-Duplex Two-Speaker Conversational Dataset
tags:
- audio
- speech
- conversational
- multi-speaker
- full-duplex
- turn-taking
- english
size_categories:
- n<1K
gated: true
extra_gated_prompt: >-
  This dataset contains recordings of identifiable human speakers and
  self-reported demographic attributes. By requesting access you agree to use
  the data for research and AI development only, to follow the license terms,
  and to make no attempt to identify or contact any speaker.
extra_gated_fields:
  Full name: text
  Affiliation: text
  Intended use: text
  I agree to the data use terms: checkbox
---

# American English Full-Duplex Two-Speaker Conversational Dataset

## Dataset summary

Natural, **unscripted, two-speaker English conversations** recorded by fluent
English speakers based in the **United States and Canada**. Each session is a
~15-minute spontaneous discussion between a matched pair of speakers on everyday
topics — personal experiences, hobbies, workplace challenges, or opinions that
have changed over time.

The recordings are designed to support the development of next-generation AI
systems — helping them better understand **natural speech patterns,
conversational flow, turn-taking, and real-world human interaction**. Each
speaker is captured on an **independent, isolated audio track**, enabling
per-speaker analysis, diarization, full-duplex modeling, ASR, and TTS.

## Dataset statistics

| | |
|---|---|
| Conversations | 954 |
| Per-speaker tracks (rows) | 1,908 |
| Conversation audio | **~230 hours** |
| Isolated per-speaker audio (both tracks) | **~455 hours** |
| Avg / median conversation length | 14.5 / 15.0 min |
| Length range | 0.4 – 15.3 min |

Each conversation contributes two simultaneous isolated speaker tracks, so the
~230 hours of conversation yields ~455 hours of clean single-speaker audio.

## How it was collected

Participants were fluent U.S./Canada-based English speakers recording remotely:

- **Conversational recording** — two matched partners hold a ~15-minute recorded
  conversation. Sessions are unscripted; partners choose a topic together.
- **Natural interaction** — speakers listen actively, respond thoughtfully, and
  build on each other's ideas in a clear, natural way.
- **Topic selection** — real-life topics (personal experiences, hobbies,
  workplace challenges, evolving opinions). The chosen prompt is stored per row.
- **Audio quality** — speakers followed guidelines to keep clear, consistent
  audio throughout each session.

Each conversation produces two simultaneous per-speaker tracks (`speaker_a` and
`speaker_b`), recorded full-duplex.

## Dataset structure

**One row per speaker track**, stacked and ordered by `room_name` so a conversation's two speakers sit adjacent. Each row is one isolated voice with its own metadata; join on `room_name` to reconstruct the conversation.
Audio is the original Opus capture in a `.opus` container.

### Fields (per row = one speaker's track)

| field | description |
|-------|-------------|
| `file_name` | this speaker's isolated audio track |
| `room_name` | conversation/session key — shared by both speakers of a conversation |
| `conversation_id` | conversation identifier |
| `role` | `SPEAKER_A` or `SPEAKER_B` |
| `speaker_id` | stable speaker identifier |
| `duration_seconds` | track duration |
| `language` | spoken language of the session (`en-US`) |
| `prompt` | the conversation topic the pair discussed |
| `gender` | self-reported |
| `city`, `country` | self-reported location |
| `ethnicity` | self-reported |
| `fluent_languages` | languages the speaker is fluent in |

Rows are ordered by `room_name` so a conversation's two speakers appear
adjacent. Join on `room_name` to reconstruct a full conversation.

## Privacy & consent

Speaker **names and emails are removed**. Demographic fields are self-reported.
Recordings were collected from consenting, compensated participants for AI
research. This is a **gated** dataset — access requires agreeing to research-only
use and no re-identification.

## Audio note

Tracks are Opus in a `opus` container. Decode with `datasets>=4.0`
(torchcodec/FFmpeg). For older stacks, losslessly rewrap to `.opus`/`.ogg`
(`ffmpeg -i in.opus -c:a copy out.opus`).

## License

Released under **CC-BY-NC-4.0** (research / non-commercial). Contact OcularAI for
other licensing.