Full-duplex English conversation datasets with each speaker on a separate channel.
otoSpeech has casual English conversation between two people, recorded with a separate channel for each speaker. Overlap, backchannels, and interruptions stay on their own tracks. The three releases differ in size, processing, and license.
| Casual | Language | Speakers | Media | License | Actions |
|---|---|---|---|---|---|
| English | 2 | Commercial | |||
English casual conversation with a separate channel for each speaker.
| |||||
| English | 2 | CC BY 4.0 | |||
Raw English conversation recorded in varied real-world conditions. Microphones and rooms differ across sessions.
| |||||
| English | 2 | CC BY 4.0 | |||
Conversations selected from the 280h release by human review, with noise reduction and speech enhancement. It also adds conversations recorded after that release.
|
Hugging FaceThree releases
otoSpeech 1,125h
Our largest English set: 1,125 hours across 3,878 sessions of casual conversation at 48 kHz, under a commercial license.
otoSpeech 280h
Raw conversation at 48 kHz, recorded in varied real-world conditions. Microphones and rooms differ across sessions. CC BY 4.0.
otoSpeech 141h
Conversations selected from the 280h release by human review, with noise reduction and speech enhancement, plus conversations recorded after that release. 44.1 kHz, CC BY 4.0.
What 280h and 141h include
- Audio
- Stereo FLAC with one speaker on each channel.
- Metadata
- Session metadata, speaker profiles, redaction intervals, and participant surveys.
- Access
- By request on Hugging Face.
Speaker profile fields
Each speaker in otoSpeech 1,125h, 280h, and 141h has these fields. The values below are made up to show the format. Fields are opt-in, so any of them can be empty.
| Field | Example | Meaning |
|---|---|---|
| speaker_id | 0123456789abcdef0123456789abcdef | Speaker identifier within the session |
| channel | 0 | Audio channel for this speaker |
| age | 34 | Age |
| name | (empty) | Handle name, currently not available |
| gender | female | Self-reported gender |
| nationality | Canada | Nationality |
| birth_country | Canada | Birth country |
| dialect | Ontario | Self-reported dialect or accent |
| first_language | English | First language |
| second_language | French | Second language |
| education_level | Bachelors | Education level |
| university | Example State University | University |
| mbti | INFP | MBTI |
| job | Teacher | Occupation |
| residence_country | United States | Residence country |
FAQ
What is full-duplex speech data?
Audio of a conversation recorded with a separate channel for each speaker, at the source. Both people can talk at once, and a single mixed track would flatten the overlap, backchannels, and interruptions that separate channels keep. fullduplex.ai, a site oto runs, explains how full-duplex voice AI is built and measured.
Which English dataset should I use?
otoSpeech 1,125h is the largest and comes with a commercial license. otoSpeech 280h keeps raw recordings from varied conditions. otoSpeech 141h is reviewed by people and cleaned.
Do the hours add up across releases?
Not for 280h and 141h. The 141-hour set draws partly on the 280-hour set.
How do I get access?
otoSpeech 280h and 141h are CC BY 4.0, and you request access on Hugging Face. For otoSpeech 1,125h, send a request from this page.
More otoSpeech data: Japanese, Korean, video, multi-speaker, and egocentric conversation, or the full catalog.