Basis Global Emotions: 525 hours of expressive speech across 108 locales
The same voice actor reads the same line according to three different emotional levels (e.g. neutral, angry, & very angry). From a neutral reading to full frustration, or from ordinary speech to a whisper, Basis Global Emotions covers a
Listen to samples · Get the dataset
What's in the release
| Audio | 525.17 hours |
| Recordings | 137,696 |
| Matched series | 54,575 |
| Unique speakers | 10,420 |
| Coverage | 108 locales, 83 languages, 90 countries |
| Format | 48 kHz, 16-bit mono WAV |
Each recording is a native speaker performing a line in a described situation, with a requested emotion or delivery style. We call this instruction a steer: happy, frustrated, apologetic, sarcastic, whispering, singing, and more.
Recordings are grouped into series. In each series, one speaker reads the same line at different intensity levels:
- Emotion series: neutral, medium, full.
- Delivery series: neutral, full.
Every reading of each selected series is included, so you can work with complete comparisons rather than isolated clips.
The same words can sound very different
How does a voice move from mild frustration to full frustration? What changes when someone sounds apologetic, relieved, or sarcastic? Can a model produce those differences while keeping the speaker's identity and words consistent?
These are the comparisons Global Emotions is designed to support. A neutral reading gives you a reference for the expressive readings beside it. Holding the speaker and script constant makes it easier to study changes in rhythm, pitch, emphasis, and delivery.
The release also includes lines with scripted non-verbal cues: a sigh, a scoff, a gasp, a laugh. In total, 12,141 recordings have a cue in the instructed script, covering 13 kinds of vocalization.
You can listen to the paired readings before downloading.
Expression across languages and regions
The release spans 83 languages, including Amharic, Assamese, Cebuano, Georgian, Javanese, Kinyarwanda, Malagasy, Oromo, Wolof, and Xhosa, alongside English, Spanish, Arabic, Japanese, and others.
Locales preserve regional distinctions within a language: Brazilian and European Portuguese, Kenyan and Tanzanian Swahili, and multiple varieties of Arabic, English, and Spanish. The dataset card breaks down hours, recordings, series, and speakers for every locale.
How we collected it
Native speakers recorded on Basis's custom consumer applications and were paid for their contributions. Each speaker received a line, a situation, and a requested steer, then recorded the readings that make up the series.
All audio was rated by native speakers from its locale. We selected series using a blend of those pairwise ratings and internal models for steer following and microphone quality, balancing across steers and genders. Selection was capped at 35 minutes per speaker and ten series per line.
Audio is provided as recorded, without post-hoc denoising, enhancement, loudness normalization, or resampling. Additionally, the open-source release does not include human-validated transcriptions; those are available at an additional cost of $22.7 per audio hour.
Contributors consented to recording and licensing for AI training. The release uses stable, salted speaker identifiers and excludes names, account identifiers, and contact details. Speaker demographics are self-reported.
Start with one locale
After obtaining access on Hugging Face, install datasets[audio] and FFmpeg, and load a locale by its code:
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"]
Group rows by series_id and order intensity as neutral, medium, full to reconstruct the paired readings. Use speaker_id when creating speaker-disjoint training and evaluation sets.
One detail matters when using the labels: steer describes the series and is also present on its neutral reading. Read it together with intensity.
Likewise, given_line is the script the speaker was asked to read, not a verified transcript. Speakers may depart from the script, and transcripts are not included in this release. The README and field guide explain the structure and metadata.
Why we're releasing it
Voice models need to handle how something is said as well as the words themselves. We want researchers and builders to be able to study expressive speech across languages, train models with more control over delivery, and evaluate whether that control holds when the speaker or locale changes.
Global Emotions complements Basis Conversations 1500: spontaneous, multi-party interaction in one release, directed and paired performances in the other. Together, they offer different ways to study the range of human speech.
Get Basis Global Emotions. Free for commercial and research use under the dataset's license. For custom collections, reach us at data@withbasis.co.
