65
stars
22
commits
Sep 6, 2026
updated
README
Open Yap 1K sample
Download the sample
Download all 16 conversations (8.9 hours)
One TAR file contains both speaker tracks, transcripts, metadata, and preview audio.
The world's largest free conversational dataset: 1000 hours of two-speaker English speech, open to labs and research teams worldwide.
With this open release, we aim to close a gap in the literature: speech recorded as it happens in real life. Fisher and Switchboard assigned partners and topics to maximize variety, but the phone network capped their audio near 4 kHz. Newer corpora record at full bandwidth but still pair strangers. Open Yap 1K assigns nothing: each speaker invites someone they already know and talks freely among friends and family.
That design choice shows in the data quality. People who know each other interrupt more, backchannel more, and leave shorter gaps between turns: the behaviour a full-duplex model has to learn.
These 1000 hours are one part of a larger licensed corpus we build with frontier labs and research teams. We release them because progress in conversational AI is slower than it needs to be - and open data is the fastest way to change that for everyone. Contact us to discuss licensing the full corpus.
The HuggingFace repository only holds a sample. The full corpus is publicly available for commercial and research use, under the Open Yap 1K Data Use Agreement. Request it at theagenticdatacompany.com/open-yap-1k.
Corpus composition
| HuggingFace | Full Corpus | |
|---|---|---|
| Size | 8.9 h, 16 conversations | 1000 h, 1602 conversations |
| Speakers | 8 | 239 |
| License | CC-BY-4.0 | Open Yap 1K Data Use Agreement |
| Access | This page | Request form |
| Audio | 48 kHz, 16-bit FLAC, one file per speaker | 48 kHz, 16-bit PCM, one file per speaker |
Contact
Website: theagenticdatacompany.com
Email: christian@theagenticdatacompany.com
Files and loading
open-yap-1k/
├── README.md
├── LICENSE.txt
├── metadata.jsonl
├── open-yap-1k-sample.tar
├── preview/
│ ├── metadata.jsonl
│ └── <key>.mp3
└── shard-00000.tar to shard-00001.tar
├── <key>.preview.mp3
├── <key>.a.flac
├── <key>.b.flac
└── <key>.json
Load both speaker tracks and the full conversation record in Python:
from datasets import load_dataset
ds = load_dataset(
"webdataset",
data_files={"train": "hf://datasets/TheAgenticDataCompany/open-yap-1k/shard-*.tar"},
split="train",
)
sample = ds[0]
a, b = sample["a.flac"]["array"], sample["b.flac"]["array"]
assert len(a) == len(b) # one timeline, by construction
record = sample["json"] # metadata and both transcripts
Or fetch the files and read them with any tar reader:
hf download TheAgenticDataCompany/open-yap-1k --repo-type dataset --include "shard-*.tar" --local-dir open-yap-1k
Collection method and quality assurance
We collect the audio with our own app, which works like a phone call. A speaker invites someone they know, and the two of them simply talk. This gives us natural conversation and a straightforward way to scale collection. The trade-off is less control over the recording environment, so real-world noise appears more often. We keep the subtle noise, because it teaches a model robustness. However, we remove the long tail of noisy recordings with several ML models combined into one pipeline.
| Other datasets | Open Yap 1K | |
|---|---|---|
| Speakers | Strangers, paired by the collector | Friends and family, self-paired |
| Setting | A treated room | Real rooms, on their own devices |
| Conversation | An assigned topic | Free talk, any topic |
| What you hear | Clean turns, little overlap | Overlap, quick turns, backchannel, laughter |
| Background | Pristine | Slightly noisier, kept on purpose |
Consent and privacy
- All audio was recorded on our own platform.
- Speakers register, give explicit consent before their first recording, and are paid for their time.
- Demographics are self-reported at registration, before any recording, and are never inferred from audio.
- Speaker identifiers are pseudonymous and stable within the release. Names, contact details and account identifiers are excluded.
Known limitations
- Hand-picked, not a random draw. 16 friends; nothing about this sample's distribution generalises to the corpus.
- Transcripts are machine-generated (Deepgram Nova-3, word-level) and not human-verified. Expect errors on overlapping speech and names.
- 8 speakers. Several appear in more than one conversation, so this is a handful of voices, not a population.
- 8 of 32 tracks carry no energy above 8 kHz. Those
microphones were Bluetooth headsets, but the file is still 48 kHz; read
effective_bandwidth_hzper track before assuming full bandwidth.
Metadata
| File name | Field | Meaning |
|---|---|---|
<key>.json | conversation_id | stable pseudonym |
language | BCP-47 | |
relationship | self-reported | |
duration_seconds | conversation length | |
topics[] | subject tags | |
turns_per_minute | speaker handoffs per minute, excluding short backchannels | |
turn_taking_gap_ms | median gap between turns | |
speech_dominance | speaker A's share of spoken words, 0-1 (0.5 = equal) | |
sample_rate | 48000 Hz, as delivered | |
speakers.{a,b} | per-speaker metadata | |
transcripts.{a,b} | per-speaker transcript | |
speakers.{a,b} | speaker_id | stable pseudonym |
age_range | bucketed, e.g. "25-34" | |
gender | self-reported | |
country | ISO-3166 alpha-2, self-reported | |
education_level | self-reported | |
native_language | BCP-47, self-reported | |
accent | derived from self-reported demographics, not measured from audio | |
recording | track format and measured levels | |
avg_wpm | words per minute | |
headphones | exact on mobile, inferred on web | |
echo_cancellation | applied at capture | |
device | capture device or audio route | |
audio_metrics | measured from this track, at the rate this archive ships | |
transcripts.{a,b} | conversation_id | stable pseudonym |
speaker_index | "a" | "b" | |
language | BCP-47 | |
text | full transcript | |
words[] | { word, start, end, type } | |
corrections_applied | a reviewer edited the ASR output |
Audio is delivered un-normalised. Integrated loudness and true peak are measured and reported so a target level can be applied without probing every file.
Citation
@misc{openyap1k,
title = {Open Yap 1K: Channel-Separated English Natural Two-Speaker Conversations},
author = {The Agentic Data Company},
year = {2026},
version = {1.0},
publisher = {The Agentic Data Company},
url = {https://theagenticdatacompany.com/open-yap-1k},
}
Contributors
22 commits
Similar repos
65
stars
22
commits
Sep 6, 2026
updated
README
Open Yap 1K sample
Download the sample
Download all 16 conversations (8.9 hours)
One TAR file contains both speaker tracks, transcripts, metadata, and preview audio.
The world's largest free conversational dataset: 1000 hours of two-speaker English speech, open to labs and research teams worldwide.
With this open release, we aim to close a gap in the literature: speech recorded as it happens in real life. Fisher and Switchboard assigned partners and topics to maximize variety, but the phone network capped their audio near 4 kHz. Newer corpora record at full bandwidth but still pair strangers. Open Yap 1K assigns nothing: each speaker invites someone they already know and talks freely among friends and family.
That design choice shows in the data quality. People who know each other interrupt more, backchannel more, and leave shorter gaps between turns: the behaviour a full-duplex model has to learn.
These 1000 hours are one part of a larger licensed corpus we build with frontier labs and research teams. We release them because progress in conversational AI is slower than it needs to be - and open data is the fastest way to change that for everyone. Contact us to discuss licensing the full corpus.
The HuggingFace repository only holds a sample. The full corpus is publicly available for commercial and research use, under the Open Yap 1K Data Use Agreement. Request it at theagenticdatacompany.com/open-yap-1k.
Corpus composition
| HuggingFace | Full Corpus | |
|---|---|---|
| Size | 8.9 h, 16 conversations | 1000 h, 1602 conversations |
| Speakers | 8 | 239 |
| License | CC-BY-4.0 | Open Yap 1K Data Use Agreement |
| Access | This page | Request form |
| Audio | 48 kHz, 16-bit FLAC, one file per speaker | 48 kHz, 16-bit PCM, one file per speaker |
Contact
Website: theagenticdatacompany.com
Email: christian@theagenticdatacompany.com
Files and loading
open-yap-1k/
├── README.md
├── LICENSE.txt
├── metadata.jsonl
├── open-yap-1k-sample.tar
├── preview/
│ ├── metadata.jsonl
│ └── <key>.mp3
└── shard-00000.tar to shard-00001.tar
├── <key>.preview.mp3
├── <key>.a.flac
├── <key>.b.flac
└── <key>.json
Load both speaker tracks and the full conversation record in Python:
from datasets import load_dataset
ds = load_dataset(
"webdataset",
data_files={"train": "hf://datasets/TheAgenticDataCompany/open-yap-1k/shard-*.tar"},
split="train",
)
sample = ds[0]
a, b = sample["a.flac"]["array"], sample["b.flac"]["array"]
assert len(a) == len(b) # one timeline, by construction
record = sample["json"] # metadata and both transcripts
Or fetch the files and read them with any tar reader:
hf download TheAgenticDataCompany/open-yap-1k --repo-type dataset --include "shard-*.tar" --local-dir open-yap-1k
Collection method and quality assurance
We collect the audio with our own app, which works like a phone call. A speaker invites someone they know, and the two of them simply talk. This gives us natural conversation and a straightforward way to scale collection. The trade-off is less control over the recording environment, so real-world noise appears more often. We keep the subtle noise, because it teaches a model robustness. However, we remove the long tail of noisy recordings with several ML models combined into one pipeline.
| Other datasets | Open Yap 1K | |
|---|---|---|
| Speakers | Strangers, paired by the collector | Friends and family, self-paired |
| Setting | A treated room | Real rooms, on their own devices |
| Conversation | An assigned topic | Free talk, any topic |
| What you hear | Clean turns, little overlap | Overlap, quick turns, backchannel, laughter |
| Background | Pristine | Slightly noisier, kept on purpose |
Consent and privacy
- All audio was recorded on our own platform.
- Speakers register, give explicit consent before their first recording, and are paid for their time.
- Demographics are self-reported at registration, before any recording, and are never inferred from audio.
- Speaker identifiers are pseudonymous and stable within the release. Names, contact details and account identifiers are excluded.
Known limitations
- Hand-picked, not a random draw. 16 friends; nothing about this sample's distribution generalises to the corpus.
- Transcripts are machine-generated (Deepgram Nova-3, word-level) and not human-verified. Expect errors on overlapping speech and names.
- 8 speakers. Several appear in more than one conversation, so this is a handful of voices, not a population.
- 8 of 32 tracks carry no energy above 8 kHz. Those
microphones were Bluetooth headsets, but the file is still 48 kHz; read
effective_bandwidth_hzper track before assuming full bandwidth.
Metadata
| File name | Field | Meaning |
|---|---|---|
<key>.json | conversation_id | stable pseudonym |
language | BCP-47 | |
relationship | self-reported | |
duration_seconds | conversation length | |
topics[] | subject tags | |
turns_per_minute | speaker handoffs per minute, excluding short backchannels | |
turn_taking_gap_ms | median gap between turns | |
speech_dominance | speaker A's share of spoken words, 0-1 (0.5 = equal) | |
sample_rate | 48000 Hz, as delivered | |
speakers.{a,b} | per-speaker metadata | |
transcripts.{a,b} | per-speaker transcript | |
speakers.{a,b} | speaker_id | stable pseudonym |
age_range | bucketed, e.g. "25-34" | |
gender | self-reported | |
country | ISO-3166 alpha-2, self-reported | |
education_level | self-reported | |
native_language | BCP-47, self-reported | |
accent | derived from self-reported demographics, not measured from audio | |
recording | track format and measured levels | |
avg_wpm | words per minute | |
headphones | exact on mobile, inferred on web | |
echo_cancellation | applied at capture | |
device | capture device or audio route | |
audio_metrics | measured from this track, at the rate this archive ships | |
transcripts.{a,b} | conversation_id | stable pseudonym |
speaker_index | "a" | "b" | |
language | BCP-47 | |
text | full transcript | |
words[] | { word, start, end, type } | |
corrections_applied | a reviewer edited the ASR output |
Audio is delivered un-normalised. Integrated loudness and true peak are measured and reported so a target level can be applied without probing every file.
Citation
@misc{openyap1k,
title = {Open Yap 1K: Channel-Separated English Natural Two-Speaker Conversations},
author = {The Agentic Data Company},
year = {2026},
version = {1.0},
publisher = {The Agentic Data Company},
url = {https://theagenticdatacompany.com/open-yap-1k},
}
Similar repos
Contributors
22 commits