This dataset contains recordings of real people. By requesting access you agree not to try to identify, locate or contact any speaker, not to use the data to impersonate a specific person, not to reconstruct anything that was removed, and to use it under the terms in the LICENSE file.

Basis Global Emotions

Listen first: speech samples

Overview

525.17 hours · 137,696 recordings · 54,575 series · 10,420 speakers · 108 locales

Each recording is one native speaker performing a line in a described situation according to a steer, which is an emotion (e.g. happy, frustrated) or a direction (e.g., whispering, screaming). Recordings are organized into what we call series: the same speaker reads the same line multiple times to control for speaker identity and lexical content. An emotion series has three readings (neutral, medium, full); a delivery series has two (neutral, full). Every reading of every series is included; series_id groups them and intensity orders them.

Languages

Locale	Language	Country	Hours	Recordings	Series	Speakers
pt-BR	Portuguese	Brazil	9.97	2,382	946	80
ja-JP	Japanese	Japan	8.84	2,322	939	73
pl-PL	Polish	Poland	8.49	2,168	857	88
en-US	English	United States	8.46	2,497	995	209
uk-UA	Ukrainian	Ukraine	8.30	1,938	778	88
it-IT	Italian	Italy	8.13	2,042	805	71
es-CL	Spanish	Chile	8.06	2,199	888	81
am-ET	Amharic	Ethiopia	7.89	2,250	915	118
bg-BG	Bulgarian	Bulgaria	7.85	1,922	766	97
es-MX	Spanish	Mexico	7.73	2,092	827	86
sr-RS	Serbian	Serbia	7.67	1,838	735	86
id-ID	Indonesian	Indonesia	7.47	1,824	722	194
cs-CZ	Czech	Czechia	7.38	1,999	794	131
es-PE	Spanish	Peru	7.31	1,940	769	125
ro-RO	Romanian	Romania	7.11	1,976	793	82
mg-MG	Malagasy	Madagascar	7.01	1,860	751	129
ka-GE	Georgian	Georgia	6.93	1,899	751	159
ru-BY	Russian	Belarus	6.85	1,658	648	70
bs-BA	Bosnian	Bosnia and Herzegovina	6.81	1,710	662	112
zu-ZA	Zulu	South Africa	6.71	2,570	1,010	144
ru-KZ	Russian	Kazakhstan	6.70	1,572	612	63
el-GR	Greek	Greece	6.66	1,875	752	120
or-IN	Odia	India	6.56	1,600	636	163
es-ES	Spanish	Spain	6.54	1,671	676	128
fr-FR	French	France	6.51	1,857	755	103
mk-MK	Macedonian	North Macedonia	6.50	1,588	622	99
es-GT	Spanish	Guatemala	6.39	1,687	674	117
as-IN	Assamese	India	6.35	1,612	651	107
sk-SK	Slovak	Slovakia	6.32	1,514	587	82
ko-KR	Korean	South Korea	6.21	1,536	598	77
ar-MA	Arabic	Morocco	6.18	1,580	633	52
ceb-PH	Cebuano	Philippines	6.13	1,426	580	98
tr-TR	Turkish	Turkey	6.12	1,601	643	72
mn-MN	Mongolian	Mongolia	5.98	1,593	641	114
ml-IN	Malayalam	India	5.93	1,349	537	162
uz-UZ	Uzbek	Uzbekistan	5.91	1,381	547	143
sl-SI	Slovenian	Slovenia	5.90	1,716	677	74
ms-MY	Malay	Malaysia	5.88	1,501	593	122
lt-LT	Lithuanian	Lithuania	5.83	1,502	576	87
su-ID	Sundanese	Indonesia	5.80	1,493	591	123
ne-NP	Nepali	Nepal	5.76	1,317	536	132
sq-AL	Albanian	Albania	5.73	1,606	650	107
es-EC	Spanish	Ecuador	5.73	1,381	556	68
ar-JO	Arabic	Jordan	5.71	1,443	572	72
hr-HR	Croatian	Croatia	5.66	1,377	539	98
sw-TZ	Swahili	Tanzania	5.64	1,276	509	122
cnr-ME	Montenegrin	Montenegro	5.61	1,301	507	90
sv-SE	Swedish	Sweden	5.47	1,371	542	117
fil-PH	Filipino	Philippines	5.26	1,290	517	114
az-AZ	Azerbaijani	Azerbaijan	5.24	1,407	561	142
es-CO	Spanish	Colombia	5.16	1,333	532	82
pt-PT	Portuguese	Portugal	5.09	1,389	538	54
hu-HU	Hungarian	Hungary	5.09	1,352	544	115
mr-IN	Marathi	India	4.96	1,213	471	88
af-ZA	Afrikaans	South Africa	4.89	1,284	509	99
bn-BD	Bengali	Bangladesh	4.73	1,113	436	136
xh-ZA	Xhosa	South Africa	4.64	1,652	653	101
ar-DZ	Arabic	Algeria	4.64	1,341	528	187
kn-IN	Kannada	India	4.62	1,071	420	93
en-CA	English	Canada	4.58	1,325	507	97
bn-IN	Bengali	India	4.53	1,095	429	99
lv-LV	Latvian	Latvia	4.51	1,125	433	78
jv-ID	Javanese	Indonesia	4.30	1,205	485	176
ur-PK	Urdu	Pakistan	4.29	1,113	450	116
es-AR	Spanish	Argentina	4.27	1,106	447	70
fi-FI	Finnish	Finland	4.24	1,109	445	74
te-IN	Telugu	India	4.23	906	360	132
de-DE	German	Germany	4.17	1,159	455	50
gu-IN	Gujarati	India	4.12	1,060	417	122
en-GB	English	United Kingdom	4.11	1,022	398	58
sw-KE	Swahili	Kenya	4.10	936	367	88
ar-TN	Arabic	Tunisia	4.08	1,339	540	163
rw-RW	Kinyarwanda	Rwanda	4.02	1,077	420	132
he-IL	Hebrew	Israel	4.01	1,043	404	78
kk-KZ	Kazakh	Kazakhstan	3.92	1,047	426	54
nl-NL	Dutch	Netherlands	3.90	1,075	415	59
hy-AM	Armenian	Armenia	3.86	1,047	415	95
ar-SA	Arabic	Saudi Arabia	3.68	927	368	43
ar-LB	Arabic	Lebanon	3.66	956	391	116
si-LK	Sinhala	Sri Lanka	3.57	1,016	401	126
ar-EG	Arabic	Egypt	3.54	858	340	45
es-DO	Spanish	Dominican Republic	3.50	911	363	105
vi-VN	Vietnamese	Vietnam	3.49	1,067	411	101
ta-IN	Tamil	India	3.44	845	325	79
hi-IN	Hindi	India	3.41	857	339	47
lg-UG	Luganda	Uganda	3.37	811	315	95
ar-LY	Arabic	Libya	3.27	918	373	160
et-EE	Estonian	Estonia	3.12	774	298	60
ky-KG	Kyrgyz	Kyrgyzstan	3.11	798	307	96
th-TH	Thai	Thailand	2.98	1,260	516	189
el-CY	Greek	Cyprus	2.98	843	328	98
pa-IN	Punjabi	India	2.73	650	254	140
nb-NO	Norwegian	Norway	2.59	749	292	51
om-ET	Oromo	Ethiopia	2.10	549	206	45
da-DK	Danish	Denmark	1.60	436	171	67
lo-LA	Lao	Laos	1.55	614	245	70
yue-HK	Cantonese	Hong Kong	1.44	429	164	78
sq-XK	Albanian	Kosovo	1.30	368	147	47
zh-TW	Mandarin	Taiwan	1.28	337	130	50
ak-GH	Akan	Ghana	1.26	344	131	87
ln-CD	Lingala	DR Congo	1.16	271	103	70
km-KH	Khmer	Cambodia	1.00	418	164	36
ps-PK	Pashto	Pakistan	1.00	301	112	56
so-SO	Somali	Somalia	1.00	283	111	73
en-AU	English	Australia	1.00	275	107	31
tk-TM	Turkmen	Turkmenistan	1.00	247	95	48
en-NZ	English	New Zealand	1.00	244	92	17
wo-SN	Wolof	Senegal	0.78	294	111	47

How to Use

Each locale is a configuration named by its code. Install datasets[audio] and FFmpeg first.

from datasets import load_dataset

ds = load_dataset("basis-ai/basis-global-emotions", "ja-JP", split="train")
row = ds[0]
row["audio"], row["steer"], row["intensity"], row["given_line"]

Rows are recordings. Group them by series_id and order intensity as neutral, medium, full to get one speaker's readings of one line. Join speaker_id to speaker in speakers.jsonl for speaker-disjoint splits. For every locale at once, loop over the configurations or read data/*/*.parquet directly.

Dataset Structure

README.md                        this card
FIELDS.md                        every column and field
LICENSE                          the terms
dataset_info.json                statistics for the release and for every locale
locales.json                     every locale code with its language and country names
speakers.jsonl                   one row per speaker across the release
data/<language>-<COUNTRY>/       the locale's recordings as parquet shards: 16 columns, the audio bytes in `audio`

Each row has 16 columns: rank, series_id, intensity, recording_id, audio, speaker_id, locale, gender, age_band, steer, given_line, nonverbal_cues, situation, duration_s, nisqa_mos, nisqa_noise; audio carries the recording as recorded (WAV, 16-bit PCM, 48 kHz, mono). Field meanings: FIELDS.md.

Dataset Statistics

Measure	Value
Hours	525.17
Recordings	137,696
Series	54,575
Speakers	10,420
Locales / languages / countries	108 / 83 / 90
Hours by declared gender	men 215.8 · women 308.1 · other 1.3

Steers

Every reading of a series carries steer, the neutral reading included. Counts are series.

Steer	Series
affectionate	1,150
afraid	1,171
amused	1,626
angry	1,588
apologetic	1,574
assertive	2,440
bored	928
calm	1,236
confused	1,119
crying	1,337
curious	1,422
disappointed	1,117
disgusted	1,192
dramatic	1,899
embarrassed	1,020
encouraging	2,272
envious	947
excited	1,420
flirtatious	927
frustrated	1,360
grateful	1,109
guilty	1,034
happy	1,842
hesitant	1,723
laughing	2,890
mumbling	1,604
nostalgic	1,141
playful	2,431
proud	1,306
relieved	1,074
sad	1,441
sarcastic	1,996
screaming	264
shouting	498
singing	1,751
soothing	1,584
surprised	1,376
whispering	1,766

Non-verbal cues

Some lines ask the speaker to perform a non-verbal cue at some point in the line (e.g., sigh, scoff, tongue click, …). The cue is written inside given_line in square brackets, in the line's own language, exactly as the speaker saw it; nonverbal_cues lists the kinds of those cues in English, in order, and is empty when the line has none.

12,141 of 137,696 recordings (8.82%) have a non-verbal cue in the instructed script, across 13 cue kinds. As a reminder, given_line is what the voice actor was instructed to read, it is not a human transcription or ASR rendering.

Scripted cue	Recordings
laugh	1,642
sigh	1,349
scoff	1,317
tongue click	1,070
breathe	1,061
gasp	918
groan	900
sniffle	875
sob	826
clear throat	826
whimper	766
yawn	537
cough	62

Audio

WAV, 16-bit PCM, 48 kHz, exactly as recorded, no post-hoc denoising, enhancement, loudness normalization, resampling, etc.

Audio quality

Every recording must pass minimum NISQA requirements (MOS and background-noise scores, each 1 to 5), with the exception of whisper and laughing deliveries, which instead have their own acoustic check. The minimums are 3.80 / 3.90 in every locale except kk-KZ and ko-KR (3.75 / 3.90), zh-TW (3.76 / 3.90) and hi-IN (3.80 / 3.85).

Speakers

One line per speaker in speakers.jsonl: speaker (a stable anonymous id, consistent across the locales of this release), gender, age_band, country, region, the locales the speaker appears in, and the speaker's volume in the release (recordings, series, seconds).

Transcripts

An important note is that given_line is the text the speaker was asked to read, not a transcription of the recording: a speaker may have departed from the line. Human-validated transcripts are not included in this release; they can be requested from data@withbasis.co.

Dataset Creation

Recording

Native speakers record steered reads on Basis's custom consumer applications and are paid for their contributions. The same speaker reads the same instructed line across intensity levels or a requested delivery style.

Selection

All audio was rated by native speakers from its locale. Series were then chosen by a blend of those pairwise ratings and our internal models for steer following and microphone quality, balanced across steers and genders, with at most 35 minutes per speaker and ten series per line. rank is the order of that pick within the locale according to our blended quality scores.

Consent and Privacy

Every participant agreed at signup to be recorded and to license their recordings for AI training, including resale to third parties and training of speech recognition, text-to-speech and voice-cloning systems, and waived their right of publicity for AI training. Every line is a scripted text, so no personal details are spoken. We additionally take extra effort to remove suspected minors from the release: speakers declare their age at signup, every speaker's voice is screened by an offline model that detects underage voices, and flagged speakers were reviewed by a person before shipping. Speaker identifiers are salted and stable; no names, account identifiers or contact data are in the release. Demographics are self-reported at signup and never inferred from audio.

The license is deliberately permissive, for research and commercial use alike. The Basis Data License grants a worldwide, royalty-free, perpetual license to download, store, process, annotate, modify and combine the data, and to pretrain, train, fine-tune, distill, evaluate and benchmark models on it, for research and commercial purposes. Models and results built on it are yours: you may publish, distribute, license and sell them, and operate commercial products, hosted inference services and APIs with them, with no royalties, no revenue sharing and no further permission from Basis, and no obligation to publish code, weights or methods. Two things are not allowed: reselling or redistributing the recordings themselves (or copies and derivatives that contain them), and identifying, locating or contacting speakers, impersonating a specific person, or reconstructing removed information. See LICENSE for the full terms.

Additional Information

Verification

CHECKSUMS.sha256 lists the SHA-256 of each root file; the repository's file listing shows each parquet shard's SHA-256.

Licensing

The terms are in LICENSE. Custom licenses are available on request: data@withbasis.co.

Citation

@misc{basis2026globalemotions,
  title     = {Basis Global Emotions},
  author    = {{Basis}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/basis-ai/basis-global-emotions}
}

Downloads last month: 409

Total file size:

168 GB

Article mentioning basis-ai/basis-global-emotions

Basis Global Emotions: 525 hours of expressive speech across 108 locales