Voice AI built for African code-switching can outperform global models that treat multilingual speech as a rounding error.
Evidence For
Intron's Sahara v2.5 achieves 34.3% word error rate across 12 African languages, beating Google's Gemini 3.6 (53.8%) by nearly 20 percentage points.
The model handles code-switching (mixing languages mid-sentence) in languages like Hausa, Zulu, Swahili, and Luganda—something global labs have largely ignored.
