OSynth is released under CC BY 4.0. The people, persona names, organizations, meetings, events, and scenarios depicted in OSynth are fictional. Any resemblance or overlap with real people (living or deceased), organizations, or events is coincidental. Synthetic persona names do not identify the speakers whose recordings conditioned voice synthesis.

OSynth — audio

Rendered audio for the OSynth corpus: 2,126 meetings, 570 hours, plus a synthesized enrollment clip per speaker. Transcripts, workspace context and reference notes are in OSynth — text; join on workspace_id + meeting_id.

All audio is 16 kHz mono FLAC, stored in Parquet with the Audio feature.

Splits

test is the held-out evaluation set (15 workspaces, 238 meetings); every result in the paper is computed on it. train is everything else.

cohort	split	workspaces	meetings	hours	description
eval	test	15	238	19.8	Held-out evaluation set (1-day and 14-day workspaces); all paper results
month	train	20	530	40.1	25-day workspaces with many short meetings
week	train	27	690	207.5	1-week workspaces
day	train	39	245	53.1	1-day workspaces
scenario	train	12	116	8.8	Product-design scenario workspaces with a four-person team
long_meeting	train	47	307	240.8	2-3 day workspaces with few, long (~47 min) meetings

Configs

channels — one row per final audio channel of a meeting:

channel == "room" (kind == "room"): the conference-room microphone of an in-person or hybrid meeting, with room acoustics and background noise applied to the in-room speakers.
otherwise (kind == "speaker"): the dry track of one remote or virtual participant (speaker_id).

All channels of a meeting start at t=0 and may differ in length by a few seconds; the meeting mixdown is their sum. Each row carries the meeting's reference rttm.

enrollments — one row per speaker: a synthesized enrollment clip in that speaker's voice and its text. 8 speakers (in 7 workspaces) have no enrollment; the text dataset's meetings.enrollments_missing lists them.

from datasets import load_dataset
ch = load_dataset("iporada-otterai/osynth-audio", "channels", split="test", streaming=True)

Rows are large (a 47-minute channel is ~40 MB), so prefer streaming=True or filter by meeting_id before decoding.

Voices

Audio is synthesized using Qwen3-TTS, conditioned on short English voice prompts from Mozilla Common Voice Spontaneous Speech 3.0 (sps-corpus-3.0, CC0 1.0). Source prompt recordings are not included. Persona names are fictional and are not the identities of source speakers. In keeping with the source-data privacy notice, do not attempt to identify source speakers. This notice does not claim that the underlying voice characteristics were invented.

Planned improvements for v2

We plan to address the following in v2:

Bracketed sound descriptions. Some turns contain annotations such as [car noise] that the synthesized speaker reads aloud. These words are actually spoken, so their inclusion in the reference transcript is appropriate for WER scoring, but the result is unrealistic. We plan to handle these annotations without having speakers read them aloud.
Speaker consistency across turns. Improve the consistency of each speaker's voice from one turn to the next.
Abrupt turn endings. Address turns that are cut off abruptly by voice activity detection (VAD).

Feedback, examples of affected turns, and suggestions are welcome in the Community discussions.

Attribution

Credit the OSynth contributors, link to this dataset and the CC BY 4.0 license, and indicate any modifications. The dataset URL is the attribution reference for this release.

License and access acknowledgment

This dataset is released under CC BY 4.0; see LICENSE. Reuse, including commercial reuse, is allowed subject to the license, including attribution and indicating modifications.

The people, persona names, organizations, meetings, events, and scenarios depicted in OSynth are fictional. Any resemblance or overlap with real people (living or deceased), organizations, or events is coincidental. Synthetic persona names do not identify the speakers whose recordings conditioned voice synthesis.

Downloaders must accept the license and acknowledge the fictional nature of the dataset through the Hugging Face access form. See ACCESS_AGREEMENT.md for the acknowledgment text.

Downloads last month: -

Total file size:

49.6 GB