WAXAL: The African language speech dataset Google actually open-sourced

For years, voice tech has been a game of haves and have-nots. Siri gets English, Alexa gets Spanish, but try finding a decent ASR model for Yoruba or Twi. The gap isn’t just annoying — it’s structural. Without quality data, you can’t build anything useful.

Google Research has been quietly working on this since 2021, and now they’re dropping WAXAL into the open. Not just a small sample or a research-only release — this is a proper dataset under CC-BY-4.0, meaning anyone can use it, modify it, build on it.

What’s actually in the box

WAXAL covers 27 Sub-Saharan African languages spoken by over 100 million people across 26+ countries. That’s not everything — Africa has over 2,000 languages — but it’s a serious start. The initial release includes:

• 1,846 hours of transcribed natural speech for ASR
• 565 hours of high-fidelity recordings for TTS

What impressed me is how they collected the ASR data. Instead of having people read boring scripts (which always sounds stilted), they showed participants images from Google’s Open Images dataset and asked them to describe what they saw in their native language. This approach captures real speech patterns — tonal variations, code-switching, the way people actually talk. The resulting audio is far more natural than the typical “please read these sentences” corpus.

The TTS side took a different approach. Local community members worked in pairs, drafting scripts of 10,000–20,000 words, alternating between reading and recording. Some participants even used project funding to build custom studio boxes for professional-grade acoustics. That level of community involvement makes a difference — you get recordings that sound like real people, not robots reading in a closet.

Why this matters more than most dataset releases

I’ve seen too many “open” datasets that turn out to be restricted, require special approval, or only cover a handful of well-studied languages. WAXAL is genuinely different. The CC-BY-4.0 license means startups, researchers, and even hobbyists can use it without jumping through hoops.

This is particularly important for African AI ecosystems. Local developers and researchers have been at a massive disadvantage compared to teams working with English or Mandarin. WAXAL doesn’t solve everything, but it removes one of the biggest barriers to entry.

What’s missing (and what’s next)

27 languages is a lot, but it’s still a fraction of what’s needed. Languages like Hausa, Swahili, and Zulu are covered, but many others aren’t. Google says they intend for WAXAL to “continuously evolve and expand,” which is good — but I’ll believe it when I see the next release.

Also, 1,846 hours sounds impressive, but for ASR, more is almost always better. Large tech companies typically train on tens of thousands of hours. Still, this is a foundation, and for many of these languages, it’s the first decent dataset available.

The bottom line

WAXAL is a genuine contribution to the field. It’s not the first African language speech dataset, but it’s probably the most comprehensive openly licensed one. If you’re working on voice tech for African languages, this is your starting point. If you’re not, it’s still worth paying attention to — because the way Google structured the data collection (image-prompted ASR, community-driven TTS) is a model worth copying for other low-resource languages.

Links:

• WAXAL dataset
• Paper