WAXAL: The African Speech Dataset We’ve Been Waiting For

Google Research just released WAXAL, a speech dataset covering 27 Sub-Saharan African languages. It’s open, it’s big, and honestly, it’s about time.

Voice tech has been a game-changer for a lot of people, but if your language isn’t English, Mandarin, or Spanish, you’re probably still typing. That’s not a bug — it’s a data problem. You can’t build good speech systems without good speech data, and for most African languages, that data just didn’t exist. Until now.

What’s actually in WAXAL?

WAXAL is two datasets rolled into one. The ASR part — automatic speech recognition — has about 1,846 hours of transcribed natural speech. That’s not people reading scripts in a studio. The team used image prompts to get participants talking naturally about what they saw. Think describing a photo of a market or a wedding. The result is messy, authentic speech with all the tonal nuances and code-switching you’d expect from real conversation. That’s exactly what you need if you want a system that works in the wild, not just in a lab.

The TTS part — text-to-speech — adds another 565 hours of high-fidelity recordings. This one is more controlled. Local community members worked in pairs, drafting scripts and recording each other. Some even built custom studio boxes with project funding. The audio is clean, phonetically balanced, and ready for voice synthesis. If you want to build a natural-sounding voice for a virtual assistant or a reading app, this is your starting point.

Both datasets are released under CC-BY-4.0. That’s about as permissive as it gets. You can use it, modify it, even commercialize it, as long as you give credit.

Why this matters more than most dataset releases

I’ve seen a lot of “open” datasets that turn out to be impractical. Either the license is restrictive, or the data is too clean to be useful, or it covers only one or two languages. WAXAL doesn’t have those problems.

First, 27 languages is not trivial. We’re talking about languages like Yoruba, Hausa, Swahili, Zulu, and Amharic — collectively spoken by over 100 million people across more than 26 countries. That’s a huge chunk of the continent. And Google says they plan to keep adding more. That’s the right approach. Don’t wait for perfection; ship what you have and iterate.

Second, the data collection methodology is smarter than most. The ASR part uses image-prompted elicitation. Instead of having people read “The quick brown fox jumps over the lazy dog” in their native language — which produces stilted, unnatural speech — they showed them pictures and asked them to describe what they saw. That captures real linguistic variation. Tonal languages like Yoruba or Igbo are notoriously hard for ASR because pitch changes meaning. You need data that reflects how people actually speak, not how they read.

Third, the TTS data was collected by community members, not outsourced to some random vendor. That means the voices are representative. The accents are right. The cultural context is embedded. You can’t fake that with a contractor who doesn’t speak the language.

The catch? There’s always a catch

Let’s be real. 1,846 hours for ASR across 27 languages averages out to about 68 hours per language. That’s enough to train a decent model, but it’s not massive. For comparison, LibriSpeech has about 1,000 hours of English alone. So if you’re building for a single language, you might still need more data. WAXAL is a foundation, not a complete solution.

Also, the dataset is skewed toward certain regions and demographics. Google worked with academic partners, which means the data probably over-represents urban, educated speakers. Rural dialects and minority languages within those 27 aren’t fully covered. That’s a limitation worth acknowledging.

And then there’s the question of maintenance. Open datasets are great, but they rot if no one updates them. Google says they plan to expand WAXAL, but corporate priorities shift. I hope they follow through, because this is too important to abandon.

What this unlocks

If you’re building a voice assistant for an African language, WAXAL just saved you months of data collection. If you’re a researcher working on low-resource speech recognition, you now have a benchmark that actually represents real-world conditions. If you’re a startup trying to build a reading tutor for kids in Nigeria or Kenya, you have the raw material to train a TTS voice that sounds like a local teacher, not a robot.

This is the kind of infrastructure that makes the difference between “we could do this theoretically” and “we actually built it.” WAXAL won’t solve the digital divide overnight, but it’s a serious step in the right direction. And for once, the data is open enough that anyone can use it — not just the big labs.

The dataset is available now. Go grab it and build something.