ORÚKỌ: Benchmarking Identity Preservation in AI Systems for African Names

AI systems increasingly mediate how people are identified, from voice assistants to healthcare, banking, and government services, yet African personal names are routinely mispronounced, stripped of diacritics, or simplified by commercial speech systems. This is not a minor courtesy failure. In Yorùbá and Igbo, pitch and diacritics can distinguish words; in Hausa, special characters represent sounds with no plain-letter equivalent. Removing these marks can change meaning. WER/CER and MOS do not explicitly measure whether a generated name preserves its orthographic, phonetic, tonal, or semantic identity. We therefore make identity preservation an explicit, measurable objective for multilingual speech technology.