Subscribe

How to choose a full-duplex training dataset

日本語

Start with a free dataset whose terms allow commercial use. The database lists 10, from 1.94 to 1,000 hours. A few hours is enough to test your training pipeline on two-channel audio. Then add hours from one of the 8 datasets with a published price, from $10 to $242 per hour. In all 18, each speaker is recorded on their own channel.

Step 1: free datasets for commercial use

These 10 datasets cost nothing and allow commercial use. They range from 1.94 to 1,000 hours. Use one to check that your loader keeps the two speakers on separate channels and that training runs end to end. Under each licence is what it takes to get the files, as checked on each host on 13 November 2026.

DatasetHoursLanguagesLicence and access
Real-TurnTurk
Ata Technology Platforms
1.94 htrApache-2.0
Direct download
Perit Duplex Support Conversations (English)
Perit AI
5.1 henCC-BY-4.0
Direct download
otoSpeech-full-duplex-task-oriented-v2
otoearth
17.4 henCC-BY-4.0
Form, approved automatically
HCRC Map Task Corpus
University of Edinburgh / University of Glasgow (HCRC)
18 henCC-BY-4.0 (downloads from the corpus site)
Direct download
otoSpeech-full-duplex-task-oriented-20h
otoearth
20 henCC-BY-4.0
Form, approved automatically
Gridspace-Stanford Harper Valley
Gridspace / Stanford CS224S
23.7 henCC-BY-4.0
Direct download
otoSpeech-full-duplex-processed-141h
otoearth
141 henCC-BY-4.0
Form, approved automatically
otoSpeech-full-duplex-280h
otoearth
280 henCC-BY-4.0
Form, approved automatically
CANDOR
BetterUp Labs (Reece, Cooney et al.)
~860 henCANDOR Corpus License 2.1 (commercial use allowed, no redistribution)
Request, reviewed by a person
Open Yap 1K
The Agentic Data Company
1,000 henOpen Yap 1K Data Use Agreement (full corpus); CC-BY-4.0 (sample)
Request and data use agreement
  • By language: English 9, Turkish 1.
  • Perit Duplex Support Conversations is performed from scripts. otoSpeech-full-duplex-task-oriented-v2, HCRC Map Task Corpus and otoSpeech-full-duplex-task-oriented-20h were recorded while two people worked on a task together.
  • otoSpeech-full-duplex-processed-141h is a cleaned selection from otoSpeech-full-duplex-280h with some newer sessions, so the two overlap.
  • Open Yap 1K needs a request and a data use agreement for the full 1,000 h. Its 8.9 h open sample downloads directly.

Step 2: datasets with a published price

When you need more hours, 8 channel-separated datasets publish a price. The lowest price per hour comes first. The per-hour figure is the total divided by the hours. The total includes any membership the seller requires, which applies to the 5 rows from LDC (the Linguistic Data Consortium).

DatasetHoursTotalPer hourLanguagesHow it is sold
otoSpeech 1,125h
otoearth
1,125 h$11,250$10.00enCommercial data licence, bought online
Fisher English
LDC
1,960 h$42,000$21.43enFrom the LDC catalogue. Commercial use needs a for-profit membership
Greybeard (LDC2013S05)
LDC
590 h$37,750$63.98enFrom the LDC catalogue. Commercial use needs a for-profit membership
Global Conversational Speech
UsergyAI
305 h$19,825$65.00zh en es +8Commercial licence, by request
English (US) Conversational Speech
SpeechData.ai
1,500 h$135,000$90.00enCommercial (training and evaluation), by written quote
CALLFRIEND
LDC
~313 h$39,750$127.00en fr ar +10From the LDC catalogue. Commercial use needs a for-profit membership
Switchboard-1 (LDC97S62)
LDC
260 h$35,500$136.54enFrom the LDC catalogue. Commercial use needs a for-profit membership
HKUST Mandarin Telephone Speech, Part 1 (LDC2005S15)
LDC
149 h$36,000$241.61zhFrom the LDC catalogue. Commercial use needs a for-profit membership

What each total is made of:

  • otoSpeech 1,125h: The 1,125-hour pack, $11,250. Smaller packs start at $2,000 for 100 hours ($20 per hour).
  • Fisher English: LDC for-profit membership, Standard, 2026, $34,000; Fisher Parts 1 and 2, speech and transcripts, at LDC's reduced-license fee for members, $8,000. Non-members pay $16,000 and may use the data for research only.
  • Greybeard (LDC2013S05): LDC for-profit membership, Standard, 2026, $34,000; Greybeard at LDC's reduced-license fee for members, $3,750. Non-members pay $7,500 and may use the data for research only.
  • Global Conversational Speech: 305 hours at $65 per hour, as listed in the Hugging Face access request, $19,825.
  • English (US) Conversational Speech: 1,500 hours at $90 per hour, $135,000. Volume discounts are offered, and a purchase goes through a written quote.
  • CALLFRIEND: LDC for-profit membership, Standard, 2026, $34,000; One speech release for each of the ten sets with stated hours at LDC's reduced-license fee for members, seven at $500 and three at $750, $5,750. Non-members pay $11,500 for the ten releases ($1,000 each, $1,500 each for Korean Telephone Conversations Speech, Farsi Second Edition and Russian) and may use the data for research only. Each 1996 release of the six sets without stated hours adds $500 for members or $1,000 for non-members, and transcripts are separate releases that are not in this price.
  • Switchboard-1 (LDC97S62): LDC for-profit membership, Standard, 2026, $34,000; Switchboard-1 at LDC's reduced-license fee for members, $1,500. Non-members pay $3,000 and may use the data for research only.
  • HKUST Mandarin Telephone Speech, Part 1 (LDC2005S15): LDC for-profit membership, Standard, 2026, $34,000; HKUST Mandarin Telephone Speech and Transcript Data, Part 1, at LDC's reduced-license fee for members, $2,000. Non-members pay $4,000 for the speech and the transcripts and may use the data for research only.

LDC grants commercial rights only to for-profit members, who keep them after the membership ends. The membership is paid once a year, not per corpus, and the Subscription level is $40,000.

The LDC membership ($34,000 for one year) is counted in full on each LDC row. A buyer who takes all 5 pays it once: $55,000 for 3,272 hours, about $16.81 per hour. A company that is already a for-profit member pays only the fee: $8,000 for Fisher English ($4.08 per hour), $3,750 for Greybeard ($6.36 per hour), $5,750 for CALLFRIEND ($18.37 per hour), $1,500 for Switchboard-1 ($5.77 per hour) and $2,000 for HKUST Mandarin Telephone Speech, Part 1 ($13.42 per hour).

By language: English 7, Chinese 3, German 2, Spanish 2, French 2, Hindi 2, Japanese 2, Korean 2, 8 more in one dataset each. Fisher English (2004), Greybeard (2013), CALLFRIEND, Switchboard-1 (1993) and HKUST Mandarin Telephone Speech, Part 1 (2005) are telephone recordings. Totals are worked out from each seller's published list prices, before tax. Check the seller's page before you buy.

Other datasets and their terms

The other 12 channel-separated datasets publish no price, or the database holds none. Their terms are on each row. 8 of the 12 are otoearth sets that are available on request and publish no hours. Each dataset page links to the provider.

DatasetHoursLanguagesTerms
DuplexChat
Nakata, Saito, Saruwatari (U. Tokyo / AIST)
415k hen jaMIT (manifest only); audio rights stay with each podcast's rights holders
Basis Conversations 1500
Basis
1,502 hen es fr +19basis-data-license-1.0 (custom, gated)
SSSD (Scalable Spontaneous Speech Dataset)
Carnegie Mellon University (WAVLab)
727 henSSSD Data Use and License Agreement (Carnegie Mellon, no fee)
CALLHOME (JA / ZH / ES / DE / AR …)
LDC
~229 hja zh es +3LDC (paid, per-release)
otoSpeech Video, otoSpeech Scripted, otoSpeech Role-Play, otoSpeech Korean, otoSpeech Japanese, otoSpeech Face-to-Face, otoSpeech Egocentric, otoSpeech 3-Speaker
otoearth
not publisheden ja koOn request

Free for research only

For non-commercial research, 8 more datasets are free. 2 of them are synthetic speech.

DatasetHoursLanguagesSpeechLicence
MultiTalkPT / MultiTalkFT
CUHK MMLab
57.6k hen zhSyntheticCC-BY-NC-4.0 (PT, FT); CC-BY-SA-4.0 (Bench)
Vartalaap
Kapture CX
5,446 hhi enSyntheticCC-BY-NC-4.0 (gated, no redistribution)
Seamless Interaction
Meta FAIR
4,000+ henHumanCC-BY-NC-4.0
SmoothConv / DuplexConv
ASLP@NPU / QualiaLabs
2,100 hzhHumanCC-BY-NC-4.0
otoSpeech-full-duplex-turn-104h
otoearth
104.9 henHumanoto Speech License v1.0
CASPER
Johns Hopkins University (The SSSD Project)
102 henHumanCC-BY-NC-SA-4.0
Multi-stream Spontaneous Conversation (zh+en)
Magic Data Technology
15 hzh enHumanCC-BY-NC-ND-4.0
Japanese Duplex Conversation Dataset
MagicData
10.35 hjaHumanCC-BY-NC-ND-4.0

By language: English 6, Chinese 3, Hindi 1, Japanese 1.

Left out of these lists

  • AppTek Call-Center Dialogues (128.6 h) and DAI-ASR-I18N (42.1 h) record each speaker on their own channel but are released for evaluation, not for training.
  • AMI Meeting Corpus (100 h) and ICSI Meeting Corpus (72 h) are free under CC-BY-4.0 and have a microphone for each speaker, in a shared meeting room. The database does not tag them as channel-separated, so they are not counted here.

How many hours

There is no fixed number. Released models give a sense of scale.

  • Moshi was pre-trained on about 7 million hours of single-channel speech, then fine-tuned for full-duplex on about 2,000 hours of Fisher.
  • J-Moshi started from Moshi, was pre-trained on about 69,000 hours of Japanese spoken dialogue (J-CHAT), then fine-tuned on 344 hours of two-channel Japanese dialogue.
  • LLM-jp-Moshi-v1 used the same starting point and pre-training, then was fine-tuned on about 1,000 hours of conversation recorded over Zoom.

Read more: The data ceiling

Before you train

  • Each speaker is on their own channel. A mixed track blurs overlaps and backchannels.
  • The terms allow your use. Check the Licence and Commercial use lines on the dataset page, then read the original terms.
  • The speakers agreed to the recording. The Source line on each dataset page says Consented when the speakers were recorded with their consent. It says Not stated for Real-TurnTurk and Global Conversational Speech.

Guide: Training data for full-duplex models. All datasets, with filters: Datasets.

FAQ

Where can I get full-duplex training data for free?
10 channel-separated datasets in the database are free and allow commercial use, from Real-TurnTurk (1.94 h) to Open Yap 1K (1,000 h). Open Yap 1K needs a request and a data use agreement. The largest that needs neither is otoSpeech-full-duplex-280h (280 h). 8 more are free for non-commercial research.
Which full-duplex dataset with commercial rights has the lowest price?
The 10 free datasets in Step 1 cost nothing. Among the 8 with a published price, otoSpeech 1,125h has the lowest price per hour for a buyer with no LDC membership: $10.00, or $11,250 for 1,125 h. Fisher English is next at $21.43 per hour, with one year of LDC membership counted in. For a company that is already a for-profit member of LDC, Fisher English costs $8,000, about $4.08 per hour.
How many hours of two-channel conversation do I need?
There is no fixed number. Moshi was fine-tuned for full-duplex on about 2,000 hours of Fisher after pre-training on about 7 million hours of single-channel speech. J-Moshi and LLM-jp-Moshi-v1 both started from Moshi and pre-trained on about 69,000 hours of Japanese spoken dialogue. J-Moshi then used 344 hours of two-channel Japanese dialogue, and LLM-jp-Moshi-v1 about 1,000 hours.
Is there Japanese full-duplex training data?
9 channel-separated datasets for training include Japanese. None is free for commercial use. Free for non-commercial use: Japanese Duplex Conversation Dataset (10.35 h). With a published price: Global Conversational Speech ($65.00 per hour) and CALLFRIEND ($127.00 per hour). Available on request, with no published hours: otoSpeech Scripted, otoSpeech Role-Play and otoSpeech Japanese. With terms of their own: DuplexChat (a list of pointers to podcast audio, with rights that stay with each podcast), Basis Conversations 1500 (behind a request on Hugging Face, under a custom licence) and CALLHOME (sold by LDC release by release, with a Japanese release of about 49 hours).
Can I train a commercial model on a non-commercial dataset?
No. CC BY-NC and research-only terms rule out commercial use. Train a product on datasets whose terms allow it, and read the original terms before you start.

Updated