perit-ai/perit-duplex-support-en

Perit Duplex Support Conversations (English)

45 full-duplex customer-support conversations, 5.10 hours, recorded with one speaker per channel. Released by Perit under CC-BY-4.0.

Most conversational speech corpora give you a single mixed channel and leave the model to infer who was speaking. This set does the opposite: the two speakers are captured on separate channels, so per-channel voice activity is ground truth rather than something to be diarised. Overlaps, barge-ins and back-channels are left exactly as recorded, because for turn-taking work those moments are the signal rather than noise to be cleaned up.

Which channel holds which role is not constant. We measured the speech onset per channel across the release: 38 of 45 conversations open on the left channel and 7 open on the right. So do not assume channel 0 is the agent. The agent and customer columns name who played each role, and opening_channel, left_onset_s and right_onset_s give the measured onsets per file. Derive the mapping from those, or from per-channel VAD.

What is in it

Conversations	45
Duration	5.10 h (mean 6.8 min, range 2.4-17.9 min)
Audio	48 kHz, 16-bit PCM WAV, stereo
Channel layout	Two channels, one speaker each (role per channel varies — see below)
Language	Native US English
Speakers	8 performers in 4 recurring pairs
Licence	CC-BY-4.0

Scenarios span healthcare scheduling, billing and payments, airline and travel support, insurance claims, and consumer IT support. Agent and customer roles swap between files, so no voice is only ever an agent.

Processing

• Resampled to a single 48 kHz across the release (13 of 45 files were 44.1 kHz at source).
• Loudness-normalised to -16 LUFS (EBU R128, true peak -1.5 dBTP) with a two-pass loudnorm measurement, so files recorded in different sessions can be mixed in one training batch.
• Normalisation is applied to the stereo pair jointly, never per channel. The natural level difference between the two speakers is part of the recording and normalising each channel on its own would flatten it.
• No noise reduction, no gating, no ducking, no trimming of overlaps.

Fields

metadata.csv carries one row per file:

column	meaning
file_name	path to the WAV, relative to data/
duration_s	measured duration in seconds
speaker_pair	the recurring pair who performed it
session	the recording session the file came from
agent / customer	which performer took which role in this file
script_id	the scenario script, where one applies
scenario	short scenario label derived from the file name
language	always en-US in this release
sample_rate / channels	after processing: 48000 / 2
source_sample_rate	before processing, so resampling is auditable
source_relpath	the original session folder, for provenance
left_onset_s / right_onset_s	measured first speech onset in each channel, seconds
opening_channel	which channel speaks first, measured (left or right)

Intended use

Built for turn-taking and full-duplex work: end-of-turn prediction, barge-in and interruption detection, voice-activity projection, backchannel detection, and evaluating duplex voice agents. The per-channel layout means ground-truth speaker activity can be derived with per-channel VAD instead of diarisation.

Consent and provenance

Every performer was engaged and paid by Perit and consented in writing to the recording being published and trained on. These are performed role-plays, not real customer calls: no member of the public is recorded and there is no personal data to redact. Nothing here is scraped.

Limitations

• Scripted, not spontaneous. Performers work from scenario scripts. The turn dynamics are realistic but the speech is more fluent and less messy than genuine spontaneous conversation.
• English only, and one accent family. Native US English. It will not tell you how a model handles other locales.
• 8 voices. Speaker diversity is deliberately narrow so that turn-taking behaviour, not speaker variation, is what differs across files. Do not use it to evaluate speaker robustness.
• Studio-clean. Low noise floor and good channel separation. Add noise yourself if you need to test robustness.
• One recording was withheld from the release for channel bleed between the two speakers.
• Channel order is not a role label. See above: the opening channel varies by file. This is a property of how the sessions were recorded, not an error, but it means a pipeline that hardcodes "left = agent" will be wrong on some files.

The rest of the set

This is the open sample. The full 15.5 hours / 180 conversations in the same format is free on request at https://perit.ai/datasets — one form, link returned straight away.

Perit records this duplex format in other locales to order, including Indic languages, with the same channel separation and consent terms.

Citation

@misc{perit_duplex_support_en,
  title  = {Perit Duplex Support Conversations (English)},
  author = {Perit},
  year   = {2026},
  url    = {https://huggingface.co/datasets/perit-ai/perit-duplex-support-en},
  note   = {5.10 hours of full-duplex support conversations, one speaker per channel}
}

If you train or publish anything on this, tell us at support@perit.ai and we will point at your work from the dataset page.