Datasets

Find speech data by language, use case, and licence. Each entry links to its primary source.

94 datasets indexed60 curated for speech interaction

Search datasets

commercial use OK only94 of 94 datasets

Use case

Licence

Explore the corpus map and research notes

Within the 60 curated datasets: 20 allow commercial use, 26 are tagged for high interactivity, and 30 are from the 2024 to 2026 frontier. Always check the original licence before use. Last comprehensive review: 2026-10. Individual entries show their own verification date.

Which dataset for which job?

Rows are the 60 public STS-adjacent corpora, split into channel-separated full-duplex corpora and the rest. Columns are the training and evaluation roles each one supports. Hover a column for its definition, or click a row to jump to its card.

Hours

Quality

Ch-sep

Pretrain

SFT / Instr.

Dialog

Eval

High int.

Conv.

Emotion

Multi-lang

Comm.-OK

DuplexChatNakata, Saito, Saruwatari (U. Tokyo / AIST)ENJA

415k

—

Seamless InteractionMeta FAIREN

4.0k

48 kHz

SmoothConv / DuplexConvASLP@NPU / QualiaLabsZH

2.1k

—

Fisher EnglishLDCEN

2.0k

—

Basis Conversations 1500BasisENESFRDEIT+17

1.5k

48 kHz

Open Yap 1KThe Agentic Data CompanyEN

1.0k

48 kHz

CANDORBetterUp Labs (Reece, Cooney et al.)EN

850h

—

otoSpeech-full-duplex-280hotoearthEN

280h

48 kHz

Switchboard-1 (LDC97S62)LDCEN

260h

8 kHz

otoSpeech-full-duplex-processed-141hotoearthEN

141h

44.1 kHz

CALLHOME (JA / ZH / ES / DE / AR …)LDCJAZHESDEAR+1

120h

8 kHz

otoSpeech-full-duplex-turn-104hotoearthEN

105h

48 kHz

DAI-ASR-I18NDavid AIENZHJAKOHI+16

42h

—

otoSpeech-full-duplex-task-oriented-20hotoearthEN

20h

48 kHz

otoSpeech-full-duplex-task-oriented-v2otoearthEN

17h

48 kHz

Multi-stream Spontaneous Conversation (zh+en)Magic Data TechnologyZHEN

15h

16 kHz

Japanese Duplex Conversation DatasetMagicDataJA

10h

16 kHz

supportedpartial / implicitnot a fit

Dialog-interactive hours by language

The 28 real-recorded dialog corpora we track, across the ten most-spoken languages. Synthetic corpora such as Behavior-SD are excluded. Multilingual corpora split their hours evenly across declared languages; the rest fall into Other. Dashed rows have no public hours. Channel-separated corpora have per-speaker tracks, the audio usable for full-duplex training.

permissiveshare-alikegatednon-commercialcustom

How full-duplex ready are public corpora?

Public corpora with a known release month and hour count (1993 to 2026), by release date and size. The view opens on channel-separated corpora, the subset usable for full-duplex training. FD-ready means high interactivity or explicit overlap, turn-taking, or back-channel labels. Synthetic corpora help with behaviour supervision but do not replace real recordings. Click a dataset name for details.

FD-ready (high interactivity / FD tags)partial (dialog-interactive, medium)non-FD (pretrain / eval / SFT)channel-separatedsynthetic (TTS)

All datasets

Sort by

YODAS v3
ESPnet / WavLab (CMU)
Over 1.1 million hours of 48 kHz multi-channel YouTube speech under CC-BY-3.0, crawled from videos that do not overlap YODAS v1 or v2.
Hours
1.1M h
Languages
100+
First released
2026-09
Licence class
permissive
Use cases
multilingual, read / ASR
Licence
CC-BY-3.0
Primary source Paper
Granary
NVIDIA
Pseudo-labelled ASR and speech-translation data for 25 European languages under CC-BY-4.0: 192k hours with audio included and 451k hours of new transcripts for existing corpora.
Hours
643k h (192k h with audio)
Languages
en, de, fr, es, +21
First released
2025-05
Licence class
permissive
Use cases
multilingual, read / ASR
Licence
CC-BY-4.0
Primary source Paper
YODAS / YODAS2
WavLab (CMU)
500k hours of Creative Commons YouTube speech across 100+ languages. The larger YODAS v3 is listed separately.
Hours
500k h
Languages
100+
First released
2024-06
Licence class
permissive
Use cases
multilingual, read / ASR
Licence
CC-BY-3.0
Primary source Paper
DuplexChat
Nakata, Saito, Saruwatari (U. Tokyo / AIST)
415k hours of speaker-separated two-channel dialogue mined from public podcasts — distributed as metadata, not audio.
Hours
415k h
Languages
en, ja
First released
2026-07
Licence class
custom
Use cases
conversation, multilingual
Licence
Metadata manifest on Hugging Face; source-audio rights remain with each podcast
Primary source Paper
VoxPopuli
Meta AI
400k hours of unlabeled European Parliament speech in 23 languages + 1.8k transcribed + 17.3k hours of interpreter audio — dataset released CC0.
Hours
400k h
Languages
en, de, fr, es, +19
First released
2021-01
Licence class
permissive
Use cases
multilingual, read / ASR
Licence
CC0 (data) / CC BY-NC 4.0 (code & models)
Primary source Paper
Emilia / Emilia-Large
Amphion (CUHK-Shenzhen)
215.6k hours of in-the-wild multilingual speech across two licences: Emilia (101.7k h, non-commercial) and Emilia-YODAS (113.9k h, CC-BY-4.0).
Hours
215.6k h (Emilia-Large)
Languages
en, zh, ja, ko, de, fr
First released
2024-07
Licence class
non-commercial
Use cases
read / ASR, multilingual
Licence
CC-BY-NC-4.0 (Emilia) / CC-BY-4.0 (Emilia-YODAS)
Primary source Paper
J-CHAT
Sarulab (UTokyo) / Keio
~76,000 hours of Japanese spoken dialogue scraped from podcasts and YouTube — the first JP corpus targeted at dialogue-oriented speech-LMs.
Hours
76k h
Languages
ja
First released
2024-07
Licence class
non-commercial
Use cases
conversation, read / ASR
Licence
Research-only (JP Copyright Art. 30-4)
Primary source Paper
Libri-Light
Meta AI (FAIR)
60,000 hours of unlabelled English audiobooks — the corpus that made self-supervised speech practical.
Hours
60k h
Languages
en
First released
2019-12
Licence class
permissive
Use cases
read / ASR
Licence
MIT
Primary source Paper
MultiTalkPT / MultiTalkFT
CUHK MMLab
57.6k hours of synthetic long-form, multi-party English and Chinese full-duplex dialogue, plus a benchmark built from real recordings.
Hours
57.6k h (synthetic)
Languages
en, zh
First released
2026-09
Licence class
non-commercial
Use cases
conversation, multilingual
Licence
CC-BY-NC-4.0 (PT, FT); CC-BY-SA-4.0 (Bench)
Primary source Paper
Multilingual LibriSpeech (MLS)
Meta AI
50k hours of LibriSpeech-style audiobooks across 8 languages — the permissive multilingual pretraining baseline.
Hours
50k h
Languages
en, de, nl, fr, es, it, pt, pl
First released
2020-12
Licence class
permissive
Use cases
read / ASR, multilingual
Licence
CC-BY-4.0
Primary source Paper
Common Voice (Scripted Speech 27.0)
Mozilla
Crowdsourced multilingual read speech under CC0: 42.6k total and 29.3k validated hours in Scripted Speech 27.0 (September 2026).
Hours
42.6k h (29.3k validated)
Languages
295
First released
2019-06
Licence class
permissive
Use cases
read / ASR, multilingual
Licence
CC0-1.0
Primary source
ReazonSpeech
Reazon Holdings
35,000 hours of Japanese, pseudo-labelled from TV broadcasts with Whisper — the largest practical JP training pool.
Hours
35k h
Languages
ja
First released
2023-03
Licence class
share-alike
Use cases
read / ASR, conversation
Licence
CDLA-Sharing-1.0
Primary source
People's Speech
MLCommons
30k+ hours of diverse English speech harvested from Archive.org — the largest permissive English ASR corpus.
Hours
30k+ h
Languages
en
First released
2021-11
Licence class
share-alike
Use cases
read / ASR
Licence
CC-BY-SA-4.0 / CC-BY-4.0
Primary source Paper
GigaSpeech
SpeechColab
10k hours of English sourced from audiobooks, podcasts, and YouTube — broad acoustic coverage but research-only audio access.
Hours
10k h
Languages
en
First released
2021-06
Licence class
gated
Use cases
read / ASR, conversation
Licence
Research-only (gated, Apache-2.0 code)
Primary source Paper
WenetSpeech
WeNet community / NPU-ASLP
10,000+ labelled hours of multi-domain Mandarin scraped from YouTube and podcasts.
Hours
10k h labelled
Languages
zh
First released
2021-10
Licence class
non-commercial
Use cases
read / ASR, conversation
Licence
CC-BY-4.0 (non-commercial usage only)
Primary source Paper
VoxLingua107
Tallinn University of Technology
6,628 hours of spoken-language-identification data covering 107 languages.
Hours
6,628 h
Languages
multi
First released
2020-11
Licence class
permissive
Use cases
speaker, multilingual
Licence
CC-BY-4.0
Primary source Paper
SPGISpeech
Kensho (S&P Global)
5,000 hours of transcribed earnings calls — fully formatted, orthographic transcripts.
Hours
5,000 h
Languages
en
First released
2021-04
Licence class
gated
Use cases
conversation
Licence
Kensho user agreement (registration required)
Primary source Paper
MuST-C
FBK (Fondazione Bruno Kessler)
English TED-talk speech translated into 14 target languages — hundreds of hours per pair, the reference training corpus for ST.
Hours
~4k h
Languages
en → 14
First released
2019-06
Licence class
non-commercial
Use cases
translation, multilingual
Licence
CC-BY-NC-ND-4.0
Primary source Paper
Seamless Interaction
Meta FAIR
Over 4,000 hours of face-to-face dyadic conversation from more than 4,000 participants, with each participant's audio and video kept separate.
Hours
4,000+ h
Languages
en
First released
2025-06
Licence class
non-commercial
Use cases
conversation
Licence
CC-BY-NC-4.0
Primary source Paper
CVSS
Google Research
Massively multilingual S2ST corpus — 21 source languages into English, with synthetic target speech — CC-BY-4.0.
Hours
3,809 h
Languages
21 → en
First released
2022-01
Licence class
permissive
Use cases
translation, multilingual
Licence
CC-BY-4.0
Primary source Paper
IndicVoices (conversational slice)
AI4Bharat (IIT Madras)
~3,500 hours of conversational speech across all 22 scheduled Indic languages — the single largest permissive source for Hindi/Bengali/Tamil/Telugu dialogue.
Hours
~3,500 h (conv.)
Languages
hi, bn, ta, te, mr, gu, kn, ml, pa, or, as, ur, ne, sd, ks, kok, brx, doi, mai, mni, sa, sat
First released
2024-03
Licence class
permissive
Use cases
conversation, multilingual
Licence
CC-BY-4.0
Primary source Paper
CoVoST-2
Meta AI
2,880 hours of multilingual S2T translation built on top of Common Voice — the standard supervised S2ST base.
Hours
2,880 h
Languages
21 + en
First released
2020-07
Licence class
permissive
Use cases
translation, multilingual
Licence
CC0-1.0
Primary source Paper
VoxCeleb2
VGG, University of Oxford
6,112 speakers and over a million utterances — the scale tier of speaker recognition.
Hours
2,442 h
Languages
multi
First released
2018-06
Licence class
share-alike
Use cases
speaker, conversation
Licence
CC-BY-SA-4.0 (annotations; source video rights retained)
Primary source Paper
Behavior-SD
Seoul National University (NAACL 2025)
108K LLM-synthesised full-duplex dialogues (2,164 h) with explicit backchannel / interruption / filler labels — the largest publicly-downloadable FD-labeled corpus.
Hours
2,164 h
Languages
en
First released
2025-04
Licence class
permissive
Use cases
dialog, conversation
Licence
CC-BY-4.0
Primary source Paper
SmoothConv / DuplexConv
ASLP@NPU / QualiaLabs
2,100 hours of Mandarin multi-party conversation with a synchronized track per participant: 100 hours human-annotated, 2,000 hours LLM-annotated.
Hours
2,100 h (100 h + 2,000 h)
Languages
zh
First released
2026-06
Licence class
non-commercial
Use cases
conversation
Licence
CC-BY-NC-4.0
Primary source Paper
Fisher English
LDC
2,000 hours of 10-minute telephone conversations — a classic for conversational ASR and turn-taking baselines.
Hours
2,000 h
Languages
en
First released
2004-12
Licence class
gated
Use cases
conversation
Licence
LDC (paid)
Primary source
Vartalaap
Kapture CX
1,955 hours of synthetic Hindi-English customer-support calls with one stereo channel per role, built for full-duplex speech-to-speech training.
Hours
1,955 h (synthetic)
Languages
hi, en
First released
2026-08
Licence class
non-commercial
Use cases
conversation, dialog
Licence
CC-BY-NC-4.0 (gated, no redistribution)
Primary source
MLC-SLM (Interspeech 2025)
Nexdata / Interspeech 2025 challenge
1,604 hours of two-speaker conversational speech in 11 languages — the first genuinely multilingual public conversational Speech-LM benchmark.
Hours
1,604 h
Languages
en, fr, de, it, pt, es, ja, ko, ru, th, vi
First released
2025-09
Licence class
gated
Use cases
conversation, multilingual
Licence
Challenge access (CC-BY-SA-4.0 for eval)
Primary source Paper
Basis Conversations 1500
Basis
1,502 hours of multi-party conversation in 22 languages, with every speaker on a separate 48 kHz track.
Hours
1,502 h
Languages
en, es, fr, de, it, nl, pl, pt, ru, uk, tr, ar, he, hi, ja, ko, zh, ka, kk, xmf, xh, zu
First released
2026-09
Licence class
gated
Use cases
conversation, multilingual
Licence
basis-data-license-1.0 (custom, gated)
Primary source
FLEURS
Google Research
Few-shot Learning Evaluation of Universal Representations of Speech — 102 languages, ~10 h each, built on FLoRes translation pairs.
Hours
~1k h
Languages
102+
First released
2022-05
Licence class
permissive
Use cases
multilingual, translation
Licence
CC-BY-4.0
Primary source Paper
Open Yap 1K
The Agentic Data Company
1,000 hours of two-speaker English conversation between people who already know each other, one 48 kHz track per speaker, under a data use agreement.
Hours
1,000 h (8.9 h open sample)
Languages
en
First released
2026-09
Licence class
gated
Use cases
conversation
Licence
Open Yap 1K Data Use Agreement (full corpus); CC-BY-4.0 (sample)
Primary source
AISHELL-2
Beijing Shell Shell Technology
1,000 hours of clean read Mandarin, free to academia on application.
Hours
1,000 h
Languages
zh
First released
2018-08
Licence class
gated
Use cases
read / ASR
Licence
Academic research licence (application required)
Primary source Paper
KsponSpeech
ETRI / Korean AI Hub
969 hours of Korean spontaneous dyadic conversation — the canonical Korean analogue of Fisher / Switchboard.
Hours
969 h
Languages
ko
First released
2020-03
Licence class
gated
Use cases
conversation
Licence
AI Hub (registration)
Primary source Paper
LibriSpeech
Vassil Panayotov et al.
960 hours of read English audiobooks — the canonical ASR benchmark since 2015.
Hours
960 h
Languages
en
First released
2015-04
Licence class
permissive
Use cases
read / ASR
Licence
CC-BY-4.0
Primary source Paper
CANDOR
BetterUp Labs (Reece, Cooney et al.)
850 hours of naturalistic video-chat conversation across 1,656 dyads — the go-to open corpus for full-duplex turn-taking research.
Hours
850 h
Languages
en
First released
2023-03
Licence class
non-commercial
Use cases
conversation
Licence
CC-BY-NC-4.0
Primary source Paper
Multilingual TEDx
Johns Hopkins / CMU
760 hours of TEDx talks in 8 languages with speech-translation pairs.
Hours
760 h
Languages
es, fr, pt, it, ru, el, ar, de
First released
2021-02
Licence class
non-commercial
Use cases
multilingual, translation
Licence
CC-BY-NC-ND-4.0
Primary source Paper
Corpus of Spontaneous Japanese (CSJ)
NINJAL / NICT / Tokyo Tech
661 hours of Japanese spontaneous speech — the canonical academic JP corpus for prosody, POS, and dependency.
Hours
661 h
Languages
ja
First released
2004-01
Licence class
gated
Use cases
conversation, read / ASR
Licence
NINJAL (tiered)
Primary source
LibriTTS
Google / Heiga Zen et al.
585 hours of LibriSpeech re-cut at 24 kHz with punctuation restored, for synthesis.
Hours
585 h
Languages
en
First released
2019-04
Licence class
permissive
Use cases
TTS, read / ASR
Licence
CC-BY-4.0
Primary source Paper
LibriTTS-R
Google
LibriTTS restored with speech enhancement — same transcripts, cleaner audio.
Hours
585 h
Languages
en
First released
2023-05
Licence class
permissive
Use cases
TTS, read / ASR
Licence
CC-BY-4.0
Primary source Paper
TED-LIUM 3
LIUM, Université du Mans
452 hours of TED talks with aligned transcripts — the standard long-form English ASR set.
Hours
452 h
Languages
en
First released
2018-05
Licence class
non-commercial
Use cases
read / ASR
Licence
CC-BY-NC-ND-3.0
Primary source Paper
MyST Children's Conversational Speech
Boulder Learning / LDC
393 hours of children talking to a virtual science tutor — the largest child-speech corpus.
Hours
393 h
Languages
en
First released
2021-06
Licence class
non-commercial
Use cases
population, conversation
Licence
CC-BY-NC-SA-4.0 (non-commercial); commercial licence available
Primary source Paper
VoxCeleb1
VGG, University of Oxford
1,251 celebrities speaking in the wild — the reference speaker-verification set.
Hours
352 h
Languages
en
First released
2017-06
Licence class
share-alike
Use cases
speaker, conversation
Licence
CC-BY-SA-4.0 (annotations; source video rights retained)
Primary source Paper
MultiDialog
IVLLab (Yonsei / Naver)
340 hours of 9k audio-visual dialogues between 12 fluent English speakers, with emotion annotations on top of TopicalChat.
Hours
340 h
Languages
en
First released
2024-06
Licence class
non-commercial
Use cases
dialog, conversation
Licence
CC-BY-NC-SA-4.0
Primary source Paper
Hi-Fi TTS
NVIDIA
292 hours of studio-grade audiobook speech from 10 speakers at 44.1 kHz.
Hours
292 h
Languages
en
First released
2021-04
Licence class
permissive
Use cases
TTS, read / ASR
Licence
CC-BY-4.0
Primary source Paper
otoSpeech-full-duplex-280h
otoearth
280 hours of channel-separated two-speaker English conversation at 48 kHz under CC-BY-4.0 — among the largest permissive, commercially-usable real-recorded FD corpora currently indexed on Hugging Face (as of 2026-04).
Hours
280 h
Languages
en
First released
2026-02
Licence class
permissive
Use cases
conversation
Licence
CC-BY-4.0
Primary source
CN-Celeb
CSLT, Tsinghua University
Chinese speaker recognition in eleven genres — deliberately harder than VoxCeleb.
Hours
274 h
Languages
zh
First released
2019-11
Licence class
non-commercial
Use cases
speaker
Licence
CC-BY-SA-4.0, no commercial usage
Primary source Paper
Switchboard-1 (LDC97S62)
LDC
260 hours of 2,400 U.S. telephone conversations — the original turn-taking corpus on which the field was built.
Hours
260 h
Languages
en
First released
1993-01
Licence class
gated
Use cases
conversation
Licence
LDC (paid)
Primary source
SpokenWOZ
Alibaba DAMO Academy
5.7K human-to-human spoken task-oriented dialogues (203K turns, 249 h) — the canonical spoken task-oriented corpus.
Hours
249 h
Languages
en
First released
2023-05
Licence class
non-commercial
Use cases
conversation, dialog
Licence
CC-BY-NC-4.0
Primary source Paper
MSP-Podcast
UT Dallas (MSP Lab)
The largest naturalistic emotional speech corpus — podcast audio, continuously growing.
Hours
~230 h
Languages
en
First released
2017-01
Licence class
gated
Use cases
emotion, conversation
Licence
Academic licence (application required)
Primary source
MM-F2F
Xiamen University
210 hours of multi-modal face-to-face conversation with turn-taking and backchannel labels at word level.
Hours
210 h
Languages
en
First released
2025-07
Licence class
custom
Use cases
conversation, dialog
Licence
Research (see repo)
Primary source Paper
NaijaS2ST
Maltais et al.
Parallel speech-to-speech translation across four Nigerian languages plus accented English.
Hours
~200 h
Languages
ig, ha, yo, pcm, en
First released
2026-04
Licence class
custom
Use cases
translation, multilingual
Licence
Not stated in the paper — verify before use
Primary source Paper
MagicData-RAMC
Magic Data Technology / OpenSLR SLR-123
180 hours of Mandarin spontaneous two-party conversation — the canonical CN-side analogue of Switchboard, free for academic use.
Hours
180 h
Languages
zh
First released
2022-03
Licence class
non-commercial
Use cases
conversation
Licence
CC-BY-NC-ND-4.0
Primary source Paper
AISHELL-1
Beijing Shell Shell Technology
178 hours of read Mandarin under Apache-2.0 — the open baseline for Chinese ASR.
Hours
178 h
Languages
zh
First released
2017-09
Licence class
permissive
Use cases
read / ASR
Licence
Apache-2.0
Primary source Paper
otoSpeech-full-duplex-processed-141h
otoearth
141 hours of human-reviewed, denoised, channel-separated English FD conversation — the processed sibling of otoSpeech-280h.
Hours
141 h
Languages
en
First released
2026-02
Licence class
permissive
Use cases
conversation
Licence
CC-BY-4.0
Primary source
CSS10
Kyubyong Park et al.
Single-speaker TTS data in 10 languages, cut from public-domain audiobooks.
Hours
140 h
Languages
de, el, es, fi, fr, hu, ja, nl, ru, zh
First released
2019-03
Licence class
permissive
Use cases
TTS, multilingual, read / ASR
Licence
Apache-2.0
Primary source Paper
CALLHOME (JA / ZH / ES / DE / AR …)
LDC
Native-speaker telephone conversations across Japanese, Mandarin, Spanish, German, Arabic, and English — the classic non-English turn-taking set.
Hours
~120 h
Languages
ja, zh, es, de, ar, en
First released
1996-01
Licence class
gated
Use cases
conversation, multilingual
Licence
LDC (paid, per-release)
Primary source
Earnings-22
Rev.com
119 hours of global-accent English earnings calls, released for accent-robustness work.
Hours
119 h
Languages
en
First released
2022-03
Licence class
share-alike
Use cases
conversation
Licence
CC-BY-SA-4.0
Primary source Paper
MUSAN
Johns Hopkins University
Music, speech and noise — the corpus behind almost every augmentation recipe.
Hours
109 h
Languages
multi
First released
2015-10
Licence class
permissive
Use cases
noise, music
Licence
CC-BY-4.0
Primary source Paper
Indic DiarBench
Sarvam AI
About 108 hours of human-annotated multi-speaker conversation across all 22 scheduled Indian languages, for joint diarization and speaker-attributed ASR.
Hours
~108 h
Languages
hi, bn, ta, te, +18
First released
2026-07
Licence class
permissive
Use cases
conversation, multilingual
Licence
CC-BY-4.0
Primary source Paper
otoSpeech-full-duplex-turn-104h
otoearth
105 hours of dyadic English conversation with hand-made turn-taking labels on every speaker channel — the training split of Sesame's TurnBench.
Hours
105 h
Languages
en
First released
2026-06
Licence class
gated
Use cases
conversation
Licence
oto Speech License v1.0
Primary source
AMI Meeting Corpus
University of Edinburgh / Idiap
100 hours of multi-party meeting recordings with close-talking mics, video, and overlap annotation — one of the few CC-BY corpora rich in overlapping speech.
Hours
100 h
Languages
en
First released
2006-06
Licence class
permissive
Use cases
conversation
Licence
CC-BY-4.0
Primary source Paper
OleSpeech-IV
Olewave / Stanford / Microsoft / PingAn / KU / CMU
English conversational speech with accents from all world regions — human-refined speaker turns and transcripts, 100-hour open subset.
Hours
100 h (open subset)
Languages
en
First released
2025-09
Licence class
non-commercial
Use cases
conversation
Licence
Non-commercial research
Primary source Paper
AISHELL-3
Beijing Shell Shell Technology
85 hours of multi-speaker Mandarin TTS data from 218 speakers, Apache-2.0.
Hours
85 h
Languages
zh
First released
2020-10
Licence class
permissive
Use cases
TTS, read / ASR
Licence
Apache-2.0
Primary source Paper
Wall Street Journal (WSJ0 / WSJ1)
LDC
Read newspaper text that defined large-vocabulary ASR benchmarking in the 1990s.
Hours
~80 h
Languages
en
First released
1993-01
Licence class
gated
Use cases
read / ASR
Licence
LDC User Agreement (paid)
Primary source
WHAM! / WHAMR!
Whisper.ai / MERL
Real ambient noise recorded in cafés and bars, paired with speech separation mixtures.
Hours
~80 h
Languages
en
First released
2019-07
Licence class
non-commercial
Use cases
noise
Licence
CC-BY-NC-4.0
Primary source Paper
ICSI Meeting Corpus
ICSI / University of Edinburgh (AMI site)
75 real research meetings recorded with close-talking and tabletop mics.
Hours
72 h
Languages
en
First released
2004-02
Licence class
permissive
Use cases
conversation
Licence
CC-BY-4.0
Primary source
VCTK Corpus
CSTR, University of Edinburgh
109 English speakers across accents — the multi-speaker and voice-cloning staple.
Hours
44 h
Languages
en
First released
2019-11
Licence class
permissive
Use cases
TTS, read / ASR
Licence
CC-BY-4.0
Primary source
DAI-ASR-I18N
David AI
42 hours of everyday two-speaker conversation in 21 languages, each speaker on a separate channel, licensed for evaluation only.
Hours
42.1 h
Languages
en, zh, ja, ko, hi, bn, mr, ta, te, th, vi, id, tl, ar, tr, ru, de, fr, es, it, pt
First released
2026-10
Licence class
gated
Use cases
conversation, multilingual
Licence
DAI-ASR-I18N Data Use Agreement v1.1 (evaluation only)
Primary source
CHiME-6
CHiME Challenge
40+ hours of real home-party dinner conversation captured over 20 parties with multi-array + close-talk mics — overlap-rich multi-party speech.
Hours
40 h
Languages
en
First released
2020-04
Licence class
non-commercial
Use cases
conversation
Licence
Research use (CHiME Challenge terms)
Primary source Paper
Expresso
Meta AI
40 hours of studio-quality expressive speech — 4 speakers, 8 read styles, 26 improvised dialogue styles including laughter and whisper.
Hours
40 h
Languages
en
First released
2023-08
Licence class
non-commercial
Use cases
emotion, read / ASR
Licence
CC-BY-NC-4.0
Primary source Paper
JVS Corpus
Takamichi Lab (University of Tokyo)
30 hours of Japanese parallel readings across 100 speakers — the staple for multi-speaker JP TTS.
Hours
30 h
Languages
ja
First released
2019-08
Licence class
custom
Use cases
read / ASR
Licence
CC-BY-SA-4.0 (with conditions)
Primary source
ESD (Emotional Speech Dataset)
NUS / HLT Lab
Parallel emotional speech in English and Mandarin from 20 speakers.
Hours
29 h
Languages
en, zh
First released
2021-05
Licence class
custom
Use cases
emotion, TTS
Licence
Research use (see repository terms)
Primary source Paper
L2-ARCTIC
Texas A&M / Iowa State
Non-native English from 24 speakers across six first languages, with error annotation.
Hours
27 h
Languages
en
First released
2018-09
Licence class
custom
Use cases
population, read / ASR
Licence
Research use (registration required)
Primary source
LJSpeech
Keith Ito
24 hours of one speaker reading public-domain books — the default single-speaker TTS set.
Hours
24 h
Languages
en
First released
2017-07
Licence class
permissive
Use cases
TTS, read / ASR
Licence
Public Domain
Primary source
Dialogs (Russian)
Shigabeev, Latyshev
20.6 hours of studio-recorded expressive Russian dialogue with per-utterance emotion labels.
Hours
20.6 h
Languages
ru
First released
2026-07
Licence class
custom
Use cases
conversation, emotion, TTS
Licence
Not stated in the paper — verify before use
Primary source Paper
DailyTalk
KAIST
20 hours of two-speaker daily conversations with emotion and dialog-act labels — staple for dialogue TTS.
Hours
20 h
Languages
en
First released
2022-07
Licence class
share-alike
Use cases
dialog
Licence
CC-BY-SA-4.0
Primary source Paper
otoSpeech-full-duplex-task-oriented-20h
otoearth
20 hours of two-speaker English conversation recorded while the pair worked on a task, with the task state, interface events and reference answers kept on the same timeline as the audio.
Hours
20 h
Languages
en
First released
2026-08
Licence class
permissive
Use cases
conversation, dialog
Licence
CC-BY-4.0
Primary source
otoSpeech-full-duplex-task-oriented-v2
otoearth
17.4 hours of two-speaker English conversation recorded during collaborative tasks, all from sessions after the 20-hour release.
Hours
17.4 h
Languages
en
First released
2026-09
Licence class
permissive
Use cases
conversation, dialog
Licence
CC-BY-4.0
Primary source
Multi-stream Spontaneous Conversation (zh+en)
Magic Data Technology
15 hours of dual-track two-speaker conversation (10 h Mandarin + 5 h English) with per-speaker audio channels — the rare public FD-native open corpus.
Hours
15 h (10 zh + 5 en)
Languages
zh, en
First released
2024-11
Licence class
non-commercial
Use cases
conversation, multilingual
Licence
CC-BY-NC-ND-4.0
Primary source Paper
IEMOCAP
USC SAIL
12 hours of scripted and improvised dyadic acting — the emotion-recognition classic.
Hours
12 h
Languages
en
First released
2008-12
Licence class
gated
Use cases
emotion, dialog
Licence
USC SAIL (request)
Primary source
MELD
DeCLaRe Lab / SUTD
Emotion recognition in multi-party conversation, cut from a sitcom script.
Hours
~12 h
Languages
en
First released
2018-10
Licence class
custom
Use cases
emotion, conversation
Licence
Not stated (repository code is GPL-3.0)
Primary source Paper
Japanese Duplex Conversation Dataset
MagicData
10.35 hours of channel-separated two-speaker Japanese conversation — tiny, but one of the only JP open corpora with per-speaker tracks.
Hours
10.35 h
Languages
ja
First released
2025-04
Licence class
non-commercial
Use cases
conversation
Licence
CC-BY-NC-ND-4.0
Primary source
JSUT
Shinnosuke Takamichi (U. Tokyo)
10 hours of single-speaker Japanese covering everyday vocabulary, recorded in an anechoic room.
Hours
10 h
Languages
ja
First released
2017-11
Licence class
custom
Use cases
TTS, read / ASR
Licence
Custom (research free; commercial terms apply, no redistribution)
Primary source Paper
SD-Eval
Amphion (CUHK-Shenzhen)
8.76 hours of curated evaluation across emotion, accent, age, and environment — non-commercial eval-only release under CC-BY-NC-4.0.
Hours
8.76 h
Languages
en
First released
2024-06
Licence class
non-commercial
Use cases
emotion, read / ASR
Licence
CC-BY-NC-4.0
Primary source Paper
PriMock57
Babylon Health
57 mock primary-care consultations released for clinical speech research.
Hours
~9 h
Languages
en
First released
2022-05
Licence class
permissive
Use cases
population, conversation
Licence
CC-BY-4.0
Primary source Paper
TIMIT
DARPA / LDC
The phonetically transcribed corpus every acoustic-model paper cited for two decades.
Hours
5.4 h
Languages
en
First released
1993-10
Licence class
gated
Use cases
read / ASR
Licence
LDC User Agreement (paid)
Primary source
CREMA-D
University of Pennsylvania
7,442 acted clips from 91 ethnically diverse actors, rated by crowdworkers.
Hours
~5 h
Languages
en
First released
2014-08
Licence class
share-alike
Use cases
emotion
Licence
Open Database License (ODbL)
Primary source
DEMAND
Inria / University of Edinburgh
Multi-channel recordings of 18 real acoustic environments.
Hours
~5 h
Languages
multi
First released
2013-06
Licence class
share-alike
Use cases
noise
Licence
CC-BY-SA-3.0
Primary source
RAVDESS
Ryerson University
24 actors performing eight emotions at two intensities, audio and video.
Hours
~3 h
Languages
en
First released
2018-04
Licence class
non-commercial
Use cases
emotion
Licence
CC-BY-NC-SA-4.0
Primary source
InstructS2S-200K
ICTNLP (LLaMA-Omni team)
200k multi-turn speech-to-speech conversations tailored for instruction-following speech models.
Hours
—
Languages
en
First released
2024-09
Licence class
non-commercial
Use cases
dialog
Licence
CC-BY-NC-4.0
Primary source Paper
VoxDialogue
VoxDialogue authors (ICLR 2025)
4,500 multi-turn spoken dialog samples × 12 acoustic attributes — probing whether spoken dialog systems catch what text can't.
Hours
—
Languages
en
First released
2025-05
Licence class
permissive
Use cases
emotion, conversation
Licence
CC-BY-4.0 (Hugging Face card metadata)
Primary source Paper
InteractSpeech
InteractSpeech authors (EMNLP 2025)
150 hours of English speech-interaction data — synthesised + filtered real-world dialogues with precise speaker timestamps for interrupts and backchannels.
Hours
150 h (audio not released)
Languages
en
First released
2025-11
Licence class
custom
Use cases
conversation, dialog
Licence
Research (see repo)
Primary source Paper
Korean full-duplex synthetic (preview)
Wi-Fi (Hugging Face)
A 100-conversation CC-BY-4.0 taster in front of a claimed 2,000-hour synthetic Korean full-duplex corpus.
Hours
preview only (2,000.5 h claimed)
Languages
ko
First released
2026-08
Licence class
permissive
Use cases
conversation
Licence
CC-BY-4.0 (preview subset)
Primary source
AffectSpeech
Qi, Zheng, Schuller, Luo, Li
Emotional speech annotated with open-vocabulary descriptions rather than a fixed label set.
Hours
—
Languages
en
First released
2026-04
Licence class
custom
Use cases
emotion, TTS
Licence
Not stated in the paper or repo — verify before use
Primary source Paper

	what it is		lang
YODAS v3 ESPnet / WavLab (CMU)	Over 1.1 million hours of 48 kHz multi-channel YouTube speech under CC-BY-3.0, crawled from videos that do not overlap YODAS v1 or v2. multilingualspeech-readsts desk	1.1M h	100+	permissive CC-BY-3.0	2026-09
Granary NVIDIA	Pseudo-labelled ASR and speech-translation data for 25 European languages under CC-BY-4.0: 192k hours with audio included and 451k hours of new transcripts for existing corpora. multilingualspeech-readsts desk	643k h (192k h with audio)	en de fr +2	permissive CC-BY-4.0	2025-05
YODAS / YODAS2 WavLab (CMU)	500k hours of Creative Commons YouTube speech across 100+ languages. The larger YODAS v3 is listed separately. multilingualspeech-readsts desk	500k h	100+	permissive CC-BY-3.0	2024-06
DuplexChat Nakata, Saito, Saruwatari (U. Tokyo / AIST)	415k hours of speaker-separated two-channel dialogue mined from public podcasts — distributed as metadata, not audio. conversationmultilingualsts desk	415k h	en ja	custom Metadata manifest on Hugging Face; source-audio rights remain with each podcast	2026-07
VoxPopuli Meta AI	400k hours of unlabeled European Parliament speech in 23 languages + 1.8k transcribed + 17.3k hours of interpreter audio — dataset released CC0. multilingualspeech-readsts desk	400k h	en de fr +2	permissive CC0 (data) / CC BY-NC 4.0 (code & models)	2021-01
Emilia / Emilia-Large Amphion (CUHK-Shenzhen)	215.6k hours of in-the-wild multilingual speech across two licences: Emilia (101.7k h, non-commercial) and Emilia-YODAS (113.9k h, CC-BY-4.0). speech-readmultilingualsts desk	215.6k h (Emilia-Large)	en zh ja +3	non-commercial CC-BY-NC-4.0 (Emilia) / CC-BY-4.0 (Emilia-YODAS)	2024-07
J-CHAT Sarulab (UTokyo) / Keio	~76,000 hours of Japanese spoken dialogue scraped from podcasts and YouTube — the first JP corpus targeted at dialogue-oriented speech-LMs. conversationspeech-readsts desk	76k h	ja	non-commercial Research-only (JP Copyright Art. 30-4)	2024-07
Libri-Light Meta AI (FAIR)	60,000 hours of unlabelled English audiobooks — the corpus that made self-supervised speech practical. speech-read	60k h	en	permissive MIT	2019-12
MultiTalkPT / MultiTalkFT CUHK MMLab	57.6k hours of synthetic long-form, multi-party English and Chinese full-duplex dialogue, plus a benchmark built from real recordings. conversationmultilingualsts desk	57.6k h (synthetic)	en zh	non-commercial CC-BY-NC-4.0 (PT, FT); CC-BY-SA-4.0 (Bench)	2026-09
Multilingual LibriSpeech (MLS) Meta AI	50k hours of LibriSpeech-style audiobooks across 8 languages — the permissive multilingual pretraining baseline. speech-readmultilingualsts desk	50k h	en de nl +5	permissive CC-BY-4.0	2020-12
Common Voice (Scripted Speech 27.0) Mozilla	Crowdsourced multilingual read speech under CC0: 42.6k total and 29.3k validated hours in Scripted Speech 27.0 (September 2026). speech-readmultilingualsts desk	42.6k h (29.3k validated)	295	permissive CC0-1.0	2019-06
ReazonSpeech Reazon Holdings	35,000 hours of Japanese, pseudo-labelled from TV broadcasts with Whisper — the largest practical JP training pool. speech-readconversationsts desk	35k h	ja	share-alike CDLA-Sharing-1.0	2023-03
People's Speech MLCommons	30k+ hours of diverse English speech harvested from Archive.org — the largest permissive English ASR corpus. speech-readsts desk	30k+ h	en	share-alike CC-BY-SA-4.0 / CC-BY-4.0	2021-11
GigaSpeech SpeechColab	10k hours of English sourced from audiobooks, podcasts, and YouTube — broad acoustic coverage but research-only audio access. speech-readconversationsts desk	10k h	en	gated Research-only (gated, Apache-2.0 code)	2021-06
WenetSpeech WeNet community / NPU-ASLP	10,000+ labelled hours of multi-domain Mandarin scraped from YouTube and podcasts. speech-readconversation	10k h labelled	zh	non-commercial CC-BY-4.0 (non-commercial usage only)	2021-10
VoxLingua107 Tallinn University of Technology	6,628 hours of spoken-language-identification data covering 107 languages. speakermultilingual	6,628 h	multi	permissive CC-BY-4.0	2020-11
SPGISpeech Kensho (S&P Global)	5,000 hours of transcribed earnings calls — fully formatted, orthographic transcripts. conversation	5,000 h	en	gated Kensho user agreement (registration required)	2021-04
MuST-C FBK (Fondazione Bruno Kessler)	English TED-talk speech translated into 14 target languages — hundreds of hours per pair, the reference training corpus for ST. s2stmultilingualsts desk	~4k h	en → 14	non-commercial CC-BY-NC-ND-4.0	2019-06
Seamless Interaction Meta FAIR	Over 4,000 hours of face-to-face dyadic conversation from more than 4,000 participants, with each participant's audio and video kept separate. conversationsts desk	4,000+ h	en	non-commercial CC-BY-NC-4.0	2025-06
CVSS Google Research	Massively multilingual S2ST corpus — 21 source languages into English, with synthetic target speech — CC-BY-4.0. s2stmultilingualsts desk	3,809 h	21 → en	permissive CC-BY-4.0	2022-01
IndicVoices (conversational slice) AI4Bharat (IIT Madras)	~3,500 hours of conversational speech across all 22 scheduled Indic languages — the single largest permissive source for Hindi/Bengali/Tamil/Telugu dialogue. conversationmultilingualsts desk	~3,500 h (conv.)	hi bn ta +19	permissive CC-BY-4.0	2024-03
CoVoST-2 Meta AI	2,880 hours of multilingual S2T translation built on top of Common Voice — the standard supervised S2ST base. s2stmultilingualsts desk	2,880 h	21 + en	permissive CC0-1.0	2020-07
VoxCeleb2 VGG, University of Oxford	6,112 speakers and over a million utterances — the scale tier of speaker recognition. speakerconversation	2,442 h	multi	share-alike CC-BY-SA-4.0 (annotations; source video rights retained)	2018-06
Behavior-SD Seoul National University (NAACL 2025)	108K LLM-synthesised full-duplex dialogues (2,164 h) with explicit backchannel / interruption / filler labels — the largest publicly-downloadable FD-labeled corpus. dialog-pairedconversationsts desk	2,164 h	en	permissive CC-BY-4.0	2025-04
SmoothConv / DuplexConv ASLP@NPU / QualiaLabs	2,100 hours of Mandarin multi-party conversation with a synchronized track per participant: 100 hours human-annotated, 2,000 hours LLM-annotated. conversationsts desk	2,100 h (100 h + 2,000 h)	zh	non-commercial CC-BY-NC-4.0	2026-06
Fisher English LDC	2,000 hours of 10-minute telephone conversations — a classic for conversational ASR and turn-taking baselines. conversationsts desk	2,000 h	en	gated LDC (paid)	2004-12
Vartalaap Kapture CX	1,955 hours of synthetic Hindi-English customer-support calls with one stereo channel per role, built for full-duplex speech-to-speech training. conversationdialog-pairedsts desk	1,955 h (synthetic)	hi en	non-commercial CC-BY-NC-4.0 (gated, no redistribution)	2026-08
MLC-SLM (Interspeech 2025) Nexdata / Interspeech 2025 challenge	1,604 hours of two-speaker conversational speech in 11 languages — the first genuinely multilingual public conversational Speech-LM benchmark. conversationmultilingualsts desk	1,604 h	en fr de +8	gated Challenge access (CC-BY-SA-4.0 for eval)	2025-09
Basis Conversations 1500 Basis	1,502 hours of multi-party conversation in 22 languages, with every speaker on a separate 48 kHz track. conversationmultilingualsts desk	1,502 h	en es fr +19	gated basis-data-license-1.0 (custom, gated)	2026-09
FLEURS Google Research	Few-shot Learning Evaluation of Universal Representations of Speech — 102 languages, ~10 h each, built on FLoRes translation pairs. multilinguals2ststs desk	~1k h	102+	permissive CC-BY-4.0	2022-05
Open Yap 1K The Agentic Data Company	1,000 hours of two-speaker English conversation between people who already know each other, one 48 kHz track per speaker, under a data use agreement. conversationsts desk	1,000 h (8.9 h open sample)	en	gated Open Yap 1K Data Use Agreement (full corpus); CC-BY-4.0 (sample)	2026-09
AISHELL-2 Beijing Shell Shell Technology	1,000 hours of clean read Mandarin, free to academia on application. speech-read	1,000 h	zh	gated Academic research licence (application required)	2018-08
KsponSpeech ETRI / Korean AI Hub	969 hours of Korean spontaneous dyadic conversation — the canonical Korean analogue of Fisher / Switchboard. conversationsts desk	969 h	ko	gated AI Hub (registration)	2020-03
LibriSpeech Vassil Panayotov et al.	960 hours of read English audiobooks — the canonical ASR benchmark since 2015. speech-readsts desk	960 h	en	permissive CC-BY-4.0	2015-04
CANDOR BetterUp Labs (Reece, Cooney et al.)	850 hours of naturalistic video-chat conversation across 1,656 dyads — the go-to open corpus for full-duplex turn-taking research. conversationsts desk	850 h	en	non-commercial CC-BY-NC-4.0	2023-03
Multilingual TEDx Johns Hopkins / CMU	760 hours of TEDx talks in 8 languages with speech-translation pairs. multilinguals2st	760 h	es fr pt +5	non-commercial CC-BY-NC-ND-4.0	2021-02
Corpus of Spontaneous Japanese (CSJ) NINJAL / NICT / Tokyo Tech	661 hours of Japanese spontaneous speech — the canonical academic JP corpus for prosody, POS, and dependency. conversationspeech-readsts desk	661 h	ja	gated NINJAL (tiered)	2004-01
LibriTTS Google / Heiga Zen et al.	585 hours of LibriSpeech re-cut at 24 kHz with punctuation restored, for synthesis. ttsspeech-read	585 h	en	permissive CC-BY-4.0	2019-04
LibriTTS-R Google	LibriTTS restored with speech enhancement — same transcripts, cleaner audio. ttsspeech-read	585 h	en	permissive CC-BY-4.0	2023-05
TED-LIUM 3 LIUM, Université du Mans	452 hours of TED talks with aligned transcripts — the standard long-form English ASR set. speech-read	452 h	en	non-commercial CC-BY-NC-ND-3.0	2018-05
MyST Children's Conversational Speech Boulder Learning / LDC	393 hours of children talking to a virtual science tutor — the largest child-speech corpus. populationconversation	393 h	en	non-commercial CC-BY-NC-SA-4.0 (non-commercial); commercial licence available	2021-06
VoxCeleb1 VGG, University of Oxford	1,251 celebrities speaking in the wild — the reference speaker-verification set. speakerconversation	352 h	en	share-alike CC-BY-SA-4.0 (annotations; source video rights retained)	2017-06
MultiDialog IVLLab (Yonsei / Naver)	340 hours of 9k audio-visual dialogues between 12 fluent English speakers, with emotion annotations on top of TopicalChat. dialog-pairedconversationsts desk	340 h	en	non-commercial CC-BY-NC-SA-4.0	2024-06
Hi-Fi TTS NVIDIA	292 hours of studio-grade audiobook speech from 10 speakers at 44.1 kHz. ttsspeech-read	292 h	en	permissive CC-BY-4.0	2021-04
otoSpeech-full-duplex-280h otoearth	280 hours of channel-separated two-speaker English conversation at 48 kHz under CC-BY-4.0 — among the largest permissive, commercially-usable real-recorded FD corpora currently indexed on Hugging Face (as of 2026-04). conversationsts desk	280 h	en	permissive CC-BY-4.0	2026-02
CN-Celeb CSLT, Tsinghua University	Chinese speaker recognition in eleven genres — deliberately harder than VoxCeleb. speaker	274 h	zh	non-commercial CC-BY-SA-4.0, no commercial usage	2019-11
Switchboard-1 (LDC97S62) LDC	260 hours of 2,400 U.S. telephone conversations — the original turn-taking corpus on which the field was built. conversationsts desk	260 h	en	gated LDC (paid)	1993-01
SpokenWOZ Alibaba DAMO Academy	5.7K human-to-human spoken task-oriented dialogues (203K turns, 249 h) — the canonical spoken task-oriented corpus. conversationdialog-pairedsts desk	249 h	en	non-commercial CC-BY-NC-4.0	2023-05
MSP-Podcast UT Dallas (MSP Lab)	The largest naturalistic emotional speech corpus — podcast audio, continuously growing. emotionconversation	~230 h	en	gated Academic licence (application required)	2017-01
MM-F2F Xiamen University	210 hours of multi-modal face-to-face conversation with turn-taking and backchannel labels at word level. conversationdialog-pairedsts desk	210 h	en	custom Research (see repo)	2025-07
NaijaS2ST Maltais et al.	Parallel speech-to-speech translation across four Nigerian languages plus accented English. s2stmultilingualsts desk	~200 h	ig ha yo +2	custom Not stated in the paper — verify before use	2026-04
MagicData-RAMC Magic Data Technology / OpenSLR SLR-123	180 hours of Mandarin spontaneous two-party conversation — the canonical CN-side analogue of Switchboard, free for academic use. conversationsts desk	180 h	zh	non-commercial CC-BY-NC-ND-4.0	2022-03
AISHELL-1 Beijing Shell Shell Technology	178 hours of read Mandarin under Apache-2.0 — the open baseline for Chinese ASR. speech-read	178 h	zh	permissive Apache-2.0	2017-09
otoSpeech-full-duplex-processed-141h otoearth	141 hours of human-reviewed, denoised, channel-separated English FD conversation — the processed sibling of otoSpeech-280h. conversationsts desk	141 h	en	permissive CC-BY-4.0	2026-02
CSS10 Kyubyong Park et al.	Single-speaker TTS data in 10 languages, cut from public-domain audiobooks. ttsmultilingualspeech-read	140 h	de el es +7	permissive Apache-2.0	2019-03
CALLHOME (JA / ZH / ES / DE / AR …) LDC	Native-speaker telephone conversations across Japanese, Mandarin, Spanish, German, Arabic, and English — the classic non-English turn-taking set. conversationmultilingualsts desk	~120 h	ja zh es +3	gated LDC (paid, per-release)	1996-01
Earnings-22 Rev.com	119 hours of global-accent English earnings calls, released for accent-robustness work. conversation	119 h	en	share-alike CC-BY-SA-4.0	2022-03
MUSAN Johns Hopkins University	Music, speech and noise — the corpus behind almost every augmentation recipe. noisemusic	109 h	multi	permissive CC-BY-4.0	2015-10
Indic DiarBench Sarvam AI	About 108 hours of human-annotated multi-speaker conversation across all 22 scheduled Indian languages, for joint diarization and speaker-attributed ASR. conversationmultilingualsts desk	~108 h	hi bn ta +2	permissive CC-BY-4.0	2026-07
otoSpeech-full-duplex-turn-104h otoearth	105 hours of dyadic English conversation with hand-made turn-taking labels on every speaker channel — the training split of Sesame's TurnBench. conversationsts desk	105 h	en	gated oto Speech License v1.0	2026-06
AMI Meeting Corpus University of Edinburgh / Idiap	100 hours of multi-party meeting recordings with close-talking mics, video, and overlap annotation — one of the few CC-BY corpora rich in overlapping speech. conversationsts desk	100 h	en	permissive CC-BY-4.0	2006-06
OleSpeech-IV Olewave / Stanford / Microsoft / PingAn / KU / CMU	English conversational speech with accents from all world regions — human-refined speaker turns and transcripts, 100-hour open subset. conversationsts desk	100 h (open subset)	en	non-commercial Non-commercial research	2025-09
AISHELL-3 Beijing Shell Shell Technology	85 hours of multi-speaker Mandarin TTS data from 218 speakers, Apache-2.0. ttsspeech-read	85 h	zh	permissive Apache-2.0	2020-10
Wall Street Journal (WSJ0 / WSJ1) LDC	Read newspaper text that defined large-vocabulary ASR benchmarking in the 1990s. speech-read	~80 h	en	gated LDC User Agreement (paid)	1993-01
WHAM! / WHAMR! Whisper.ai / MERL	Real ambient noise recorded in cafés and bars, paired with speech separation mixtures. noise	~80 h	en	non-commercial CC-BY-NC-4.0	2019-07
ICSI Meeting Corpus ICSI / University of Edinburgh (AMI site)	75 real research meetings recorded with close-talking and tabletop mics. conversation	72 h	en	permissive CC-BY-4.0	2004-02
VCTK Corpus CSTR, University of Edinburgh	109 English speakers across accents — the multi-speaker and voice-cloning staple. ttsspeech-read	44 h	en	permissive CC-BY-4.0	2019-11
DAI-ASR-I18N David AI	42 hours of everyday two-speaker conversation in 21 languages, each speaker on a separate channel, licensed for evaluation only. conversationmultilingualsts desk	42.1 h	en zh ja +18	gated DAI-ASR-I18N Data Use Agreement v1.1 (evaluation only)	2026-10
CHiME-6 CHiME Challenge	40+ hours of real home-party dinner conversation captured over 20 parties with multi-array + close-talk mics — overlap-rich multi-party speech. conversationsts desk	40 h	en	non-commercial Research use (CHiME Challenge terms)	2020-04
Expresso Meta AI	40 hours of studio-quality expressive speech — 4 speakers, 8 read styles, 26 improvised dialogue styles including laughter and whisper. emotionspeech-readsts desk	40 h	en	non-commercial CC-BY-NC-4.0	2023-08
JVS Corpus Takamichi Lab (University of Tokyo)	30 hours of Japanese parallel readings across 100 speakers — the staple for multi-speaker JP TTS. speech-readsts desk	30 h	ja	custom CC-BY-SA-4.0 (with conditions)	2019-08
ESD (Emotional Speech Dataset) NUS / HLT Lab	Parallel emotional speech in English and Mandarin from 20 speakers. emotiontts	29 h	en zh	custom Research use (see repository terms)	2021-05
L2-ARCTIC Texas A&M / Iowa State	Non-native English from 24 speakers across six first languages, with error annotation. populationspeech-read	27 h	en	custom Research use (registration required)	2018-09
LJSpeech Keith Ito	24 hours of one speaker reading public-domain books — the default single-speaker TTS set. ttsspeech-read	24 h	en	permissive Public Domain	2017-07
Dialogs (Russian) Shigabeev, Latyshev	20.6 hours of studio-recorded expressive Russian dialogue with per-utterance emotion labels. conversationemotionttssts desk	20.6 h	ru	custom Not stated in the paper — verify before use	2026-07
DailyTalk KAIST	20 hours of two-speaker daily conversations with emotion and dialog-act labels — staple for dialogue TTS. dialog-pairedsts desk	20 h	en	share-alike CC-BY-SA-4.0	2022-07
otoSpeech-full-duplex-task-oriented-20h otoearth	20 hours of two-speaker English conversation recorded while the pair worked on a task, with the task state, interface events and reference answers kept on the same timeline as the audio. conversationdialog-pairedsts desk	20 h	en	permissive CC-BY-4.0	2026-08
otoSpeech-full-duplex-task-oriented-v2 otoearth	17.4 hours of two-speaker English conversation recorded during collaborative tasks, all from sessions after the 20-hour release. conversationdialog-pairedsts desk	17.4 h	en	permissive CC-BY-4.0	2026-09
Multi-stream Spontaneous Conversation (zh+en) Magic Data Technology	15 hours of dual-track two-speaker conversation (10 h Mandarin + 5 h English) with per-speaker audio channels — the rare public FD-native open corpus. conversationmultilingualsts desk	15 h (10 zh + 5 en)	zh en	non-commercial CC-BY-NC-ND-4.0	2024-11
IEMOCAP USC SAIL	12 hours of scripted and improvised dyadic acting — the emotion-recognition classic. emotiondialog-pairedsts desk	12 h	en	gated USC SAIL (request)	2008-12
MELD DeCLaRe Lab / SUTD	Emotion recognition in multi-party conversation, cut from a sitcom script. emotionconversation	~12 h	en	custom Not stated (repository code is GPL-3.0)	2018-10
Japanese Duplex Conversation Dataset MagicData	10.35 hours of channel-separated two-speaker Japanese conversation — tiny, but one of the only JP open corpora with per-speaker tracks. conversationsts desk	10.35 h	ja	non-commercial CC-BY-NC-ND-4.0	2025-04
JSUT Shinnosuke Takamichi (U. Tokyo)	10 hours of single-speaker Japanese covering everyday vocabulary, recorded in an anechoic room. ttsspeech-read	10 h	ja	custom Custom (research free; commercial terms apply, no redistribution)	2017-11
SD-Eval Amphion (CUHK-Shenzhen)	8.76 hours of curated evaluation across emotion, accent, age, and environment — non-commercial eval-only release under CC-BY-NC-4.0. emotionspeech-readsts desk	8.76 h	en	non-commercial CC-BY-NC-4.0	2024-06
PriMock57 Babylon Health	57 mock primary-care consultations released for clinical speech research. populationconversation	~9 h	en	permissive CC-BY-4.0	2022-05
TIMIT DARPA / LDC	The phonetically transcribed corpus every acoustic-model paper cited for two decades. speech-read	5.4 h	en	gated LDC User Agreement (paid)	1993-10
CREMA-D University of Pennsylvania	7,442 acted clips from 91 ethnically diverse actors, rated by crowdworkers. emotion	~5 h	en	share-alike Open Database License (ODbL)	2014-08
DEMAND Inria / University of Edinburgh	Multi-channel recordings of 18 real acoustic environments. noise	~5 h	multi	share-alike CC-BY-SA-3.0	2013-06
RAVDESS Ryerson University	24 actors performing eight emotions at two intensities, audio and video. emotion	~3 h	en	non-commercial CC-BY-NC-SA-4.0	2018-04
InstructS2S-200K ICTNLP (LLaMA-Omni team)	200k multi-turn speech-to-speech conversations tailored for instruction-following speech models. dialog-pairedsts desk	—	en	non-commercial CC-BY-NC-4.0	2024-09
VoxDialogue VoxDialogue authors (ICLR 2025)	4,500 multi-turn spoken dialog samples × 12 acoustic attributes — probing whether spoken dialog systems catch what text can't. emotionconversationsts desk	—	en	permissive CC-BY-4.0 (Hugging Face card metadata)	2025-05
InteractSpeech InteractSpeech authors (EMNLP 2025)	150 hours of English speech-interaction data — synthesised + filtered real-world dialogues with precise speaker timestamps for interrupts and backchannels. conversationdialog-pairedsts desk	150 h (audio not released)	en	custom Research (see repo)	2025-11
Korean full-duplex synthetic (preview) Wi-Fi (Hugging Face)	A 100-conversation CC-BY-4.0 taster in front of a claimed 2,000-hour synthetic Korean full-duplex corpus. conversationsts desk	preview only (2,000.5 h claimed)	ko	permissive CC-BY-4.0 (preview subset)	2026-08
AffectSpeech Qi, Zheng, Schuller, Luo, Li	Emotional speech annotated with open-vocabulary descriptions rather than a fixed label set. emotionttssts desk	—	en	custom Not stated in the paper or repo — verify before use	2026-04

Licence strings are copied from each maintainer's terms and grouped for scanning. Re-read the original before shipping, especially for gated and custom licences. Last swept 2026-10.

What the corpus map is telling us.

Non-English interactive corpus is thin
Classic-drawer corpora outside English are overwhelmingly monologue, broadcast, or parallel-reading — few are dyadic conversational at CANDOR / AMI scale. The frontier has only started to fill the gap (J-CHAT for pretraining; MagicHub duplex sets as early dialog entries), but a permissive, large-scale, non-English conversational corpus remains unclaimed for most languages.
Commercial-friendly turn-taking is thin
The classical turn-taking corpora (Fisher, Switchboard) are gated LDC paid releases; CANDOR (850 h) is free for research but CC-BY-NC-4.0, so not directly deployable. Among the truly permissive "high interactivity" sets, AMI (CC-BY-4.0, 100 h) and otoSpeech-full-duplex-280h (CC-BY-4.0, 280 h) stand out. The rest of the frontier wave (InteractSpeech, MM-F2F, MagicHub duplex) still ships under research / custom licenses.
Scale vs license is a pareto frontier
YODAS (500k h) is CC-BY-SA, VoxPopuli (400k h audio) is CC0 (code & pretrained models CC-BY-NC-4.0), Emilia (101k h) is share-alike, People's Speech (30k h) is CC-BY-SA, Common Voice (30k+ h) is CC0. For a strictly permissive and large multilingual base, MLS (50k h, CC-BY-4.0) remains the cleanest option — with the tradeoff that it is still audiobook-flavoured and English-heavy.
frontier watch
Interactive data is finally picking up
In the 2024–2026 window, 30 FD-oriented corpora landed: J-CHAT, MultiDialog, InstructS2S-200K, SD-Eval, VoxDialogue, MM-F2F, InteractSpeech, OleSpeech-IV, the MagicHub duplex set, and otoSpeech-280h. Modalities are diversifying (audio-only → audio+video), and channel-separated stereo is becoming standard. Licenses remain the bottleneck.

Missing a dataset or a license change? Submit an entry

Which dataset for which job?

Dialog-interactive hours by language

How full-duplex ready are public corpora?

All datasets

YODAS v3

Granary

YODAS / YODAS2

DuplexChat

VoxPopuli

Emilia / Emilia-Large

J-CHAT

Libri-Light

MultiTalkPT / MultiTalkFT

Multilingual LibriSpeech (MLS)

Common Voice (Scripted Speech 27.0)

ReazonSpeech

People's Speech

GigaSpeech

WenetSpeech

VoxLingua107

SPGISpeech

MuST-C

Seamless Interaction

CVSS

IndicVoices (conversational slice)

CoVoST-2

VoxCeleb2

Behavior-SD

SmoothConv / DuplexConv

Fisher English

Vartalaap

MLC-SLM (Interspeech 2025)

Basis Conversations 1500

FLEURS

Open Yap 1K

AISHELL-2

KsponSpeech

LibriSpeech

CANDOR

Multilingual TEDx

Corpus of Spontaneous Japanese (CSJ)

LibriTTS

LibriTTS-R

TED-LIUM 3

MyST Children's Conversational Speech

VoxCeleb1

MultiDialog

Hi-Fi TTS

otoSpeech-full-duplex-280h

CN-Celeb

Switchboard-1 (LDC97S62)

SpokenWOZ

MSP-Podcast

MM-F2F

NaijaS2ST

MagicData-RAMC

AISHELL-1

otoSpeech-full-duplex-processed-141h

CSS10

CALLHOME (JA / ZH / ES / DE / AR …)

Earnings-22

MUSAN

Indic DiarBench

otoSpeech-full-duplex-turn-104h

AMI Meeting Corpus

OleSpeech-IV

AISHELL-3

Wall Street Journal (WSJ0 / WSJ1)

WHAM! / WHAMR!

ICSI Meeting Corpus

VCTK Corpus

DAI-ASR-I18N

CHiME-6

Expresso

JVS Corpus

ESD (Emotional Speech Dataset)

L2-ARCTIC

LJSpeech

Dialogs (Russian)

DailyTalk