Google Research Names WAXAL Speech Challenge Winners After 8,837 Submissions

Google Research has named the winners of its WAXAL ASR Challenge with Zindi, a competition to improve AI transcription for African languages. Team Pasketti took first place, followed by Alban Nyantudre and Abdourahamane Ide Salifou, after a contest that drew 1,462 participants from 100 countries.

From open speech data to a competition

The challenge builds on WAXAL, an open-source speech collection Google now describes as covering 32 African languages. Its initial March release covered 27 Sub-Saharan African languages and included about 1,846 hours of transcribed natural speech for automatic speech recognition, plus more than 565 hours of recordings for speech generation. The resources were released under a Creative Commons CC-BY-4.0 license.

Google Research and Zindi asked entrants to build systems that understand and transcribe spoken language, starting with Lingala and Shona. The earlier dataset release was designed around both recognition and speech generation; the contest narrowed the immediate task to recognition. Google frames the effort around an access problem: people may speak these languages fluently while finding written interfaces less practical, and many current AI systems do not understand them.

1,462Participants

Entrants came from 100 countries, including 42 African countries.

8,837Submitted solutions

Participants spent more than 24,000 hours on the problem, according to Google.

Different routes to a transcription result

Google said the top three scores were remarkably close. But the entries approached the same task differently, illustrating the range of techniques competitors used with the challenge data rather than a single prescribed solution.

• First-place Team Pasketti, represented by Roman Solovyev and enes3774 from Kazakhstan and Canada, combined eight models so their different transcription errors could be weighed against one another.
• Second-place Nyantudre of Burkina Faso had his system process the same audio clip in several ways before selecting the transcription it trusted most.
• Third-place Salifou of Niger adapted his approach to language patterns, accounting for words running together in Lingala and using surrounding context for Shona.

The work now points beyond the leaderboard

Nyantudre, a machine-learning engineer who builds speech tools for his mother tongue, Mooré, described why voice access matters to him:

Plenty of people I grew up around speak their language perfectly but cannot read or write it. That is why speech is the only realistic way for those people to use technology at all.

Google says WAXAL is intended to keep expanding to additional languages. It points to future possibilities such as checking crop prices, finding health advice, and learning by speaking to a phone in a familiar language. Those examples remain future possibilities, not announced services.