The new research resource
On October 12, Basis announced Global Emotions, a speech dataset with roughly 525 hours from 10,420 speakers across 108 locales. It groups performances so the same speaker reads the same line at different emotional or delivery intensities. Basis offers it for commercial and research use under its dataset license.
The company says contributors were paid and consented to licensing for AI training. A practical limitation: the included script is what a speaker was asked to read, not a human-verified transcript. The release does not include those verified transcriptions. This is a dataset for speech-model research and evaluation, rather than a ready-made voiceover editor.
Source: Basis: Basis Global Emotions: 525 hours of expressive speech across 108 locales ↗
Our take: give the performance a job
For a creator making an explainer, an instruction like make it emotional is hard to evaluate. Our interpretation is that a better direction describes the listener’s situation and the meaning the line should carry. A welcoming introduction, a serious correction, and a playful transition need different decisions even when they use similar words.
Imagine the line: We have one more thing to try. In a troubleshooting video, it might reassure someone who is frustrated. In a mystery story, it might introduce uncertainty. In a quick tutorial, it might simply signal the final step. Before choosing a voice or changing its delivery, decide which of those jobs the sentence is doing.
This is a creative briefing exercise you can use with an AI tool or with your own recording. Write a short note about the intended audience, what they know already, and how they should understand the next sentence. Avoid instructions that ask the delivery to compensate for unclear writing. If the explanation is confusing, revise the script first.
A small audition with fewer moving parts
Here is a suggested experiment we have not performed. Choose a voice you are authorized to use and a short script you wrote. Keep the words fixed while trying two delivery directions available in your chosen tool. If the tool has no such controls, record two versions yourself. Save the exact directions beside the outputs so you can reproduce the comparison.
Listen once for meaning and once for clarity. On the first pass, ask whether the performance supports the message. On the second, check names, numbers, pauses, and any emphasis that could change an instruction. A dramatic version that makes the next step harder to follow is not serving this particular project.
Ask a friend to listen without knowing which version you prefer. Give them a simple question such as which recording makes the instruction easiest to understand, followed by what part caused that impression. Their explanation can be more helpful than a score. It may point to a rushed clause you can edit instead of regenerating the entire narration.
If you are building a speech model
Our suggested evaluation plan is to separate performance changes from changes in words or speaker identity. Define what your model should do before picking examples, keep a record of the comparison conditions, and include situations where exaggerated delivery would be unhelpful. Use the release documentation to understand the actual labels and licensing terms rather than making assumptions from the dataset title.
For creators, the useful lesson is smaller and immediately available: you can direct a voice with a purpose and judge whether the listener receives that purpose. You do not need to train a model to start making more deliberate narration choices.
Try this, then make it yours.
Write one 30-second script and audition two authorized deliveries against a single audience-focused question.
Follow the signal.
Our reporting starts here. Practical suggestions are our analysis, and vendor performance statements are claims unless independently verified. We haven’t hands-on tested this release.
