Datasets:
GRIOT Duplex v1
Two-channel (full-duplex) synthetic call-centre conversations, built for training and evaluating full-duplex spoken dialogue models — models that listen and speak at the same time rather than taking strict turns.
Each file is one complete conversation with one speaker per channel, so overlaps, interruptions, backchannels ("mm-hm", "right") and silences are preserved exactly as they occur in time. Single-channel corpora lose this.
The repository holds two distinct corpora. They are not interchangeable and are kept in separate directories for that reason.
audio/ — long-form |
shortform/audio/ — short-form |
|
|---|---|---|
| Conversations | 2,890 | 8,667 |
| Total audio | 496.0 h | 96.8 h |
| Languages | English + Nigerian Pidgin | en, ha, ig, sw, yo |
| Median length | 10.1 min | ~40 s |
| Turns | 216,688 | — |
Format throughout: FLAC, 24 kHz, 2 channels, PCM_16.
Layout
audio/ 2,890 long-form FLAC
render_omnivoice_emotion.json timing sheet for all 2,890
data/ prior timing sheets, kept for provenance
shortform/audio/<lang>/ 8,667 short-form FLAC, 5 languages
shortform/data/<lang>/ per-shard render records
shortform/*.json ledgers and voice pools
Channel layout — channel 0 (left) is the agent, channel 1 (right) is the customer.
The timing sheet
render_omnivoice_emotion.json records, per conversation: duration, turn count,
placements — the (channel, start, duration) of every turn — plus speaker voices,
emotion labels, per-speaker code-switch level, conversation policy and hold placements.
placements is the field that makes the corpus usable for duplex training: without it
there is no way to locate a given speaker's turns inside the stereo file. Every one of
the 2,890 rows carries it, and the manifest ID set and the audio ID set match exactly in
both directions.
Long-form conversation policies: hold 1,305, converge 996, match 589.
Median inter-turn gap 60 ms.
Both channels are wideband
An earlier release of the long-form corpus applied a simulated narrowband telephone effect (300-3400 Hz band-pass plus an 8 kHz round trip) to both channels on the conversations whose scene planned a phone call. That was wrong: the agent channel is the generation target for a duplex model, and band-limiting it teaches the model to produce narrowband speech. Only a caller's incoming audio is genuinely band-limited.
Those conversations were withdrawn and replaced. Measured on the current release, the
phone_8k-planned conversations carry 41-44% of their energy below 300 Hz and roll off at
4.6-5.0 kHz - against 0.5% and 3.0 kHz for the withdrawn versions. Both channels are
wideband throughout. The planned channel is still recorded per conversation, so anyone
who wants a band-limited caller side can apply it themselves.
Intended use
Training and evaluation of full-duplex speech-to-speech models, turn-taking and backchannel prediction, overlap-robust ASR, and conversational TTS. Part of a wider effort to build full-duplex dialogue systems for African languages.
Limitations
- Synthetic speech. Prosody and disfluency are less varied than real recordings.
- Emotion labels are weak. They describe what was intended for a speaker, and where a conversation's text did not support the intended emotion the label was reset to neutral. Do not treat them as ground truth.
- Domain is customer service. It will not represent casual or domestic speech.
- Code-switching is generated, not naturally occurring.
- The short-form and long-form corpora were produced separately and differ in length, language coverage and character. Do not pool them without checking that your use case tolerates the difference.
Licence and attribution
Released under CC BY-NC 4.0 — non-commercial use only, with attribution.
The non-commercial restriction is inherited and must be passed on. Components used in producing this dataset carry non-commercial terms of their own, so any derivative you build from it must remain non-commercial.
Access
Access is gated. Requests are reviewed manually.
- Downloads last month
- 147