Evaluating and Fine-Tuning Speech Technologies on Culturally Grounded Yorùbá Proverbs Using the Òwe-Voice Proverb Speech Corpus

Modern speech technologies have advanced rapidly for African languages, yet existing benchmarks focus overwhelmingly on read, broadcast, and conversational speech, leaving the figurative and culturally grounded registers that carry a community's oral heritage largely untested. We address this gap for Yorùbá through its proverbs (òwe), a register whose meaning depends on precise tonal realisation, rare vocabulary, and shared cultural knowledge. Using Òwe-Voice, a purpose-built corpus of roughly six hours of spoken Yorùbá proverbs from twelve speakers, we systematically evaluate nine automatic speech recognition models and three text-to-speech models, measuring recognition with Word Error Rate and Character Error Rate and synthesis with Mel-Cepstral Distortion. The benchmark reveals that proverbial speech is markedly harder than conventional speech: no baseline recognition model achieved a word error rate below 0.50. We then fine-tune a representative subset of these models on in-domain data and observe substantial gains, with the word error rate of a Yorùbá-adapted Whisper model falling from 0.88 to 0.44 and that of MMS-1B from 0.62 to 0.41, so that a moderately sized fine-tuned model surpasses a much larger general-purpose one. Fine-tuning likewise improved synthesis quality, reducing Mel-Cepstral Distortion from 16.24 to 11.31 dB. These results show that in-domain adaptation, rather than scale alone, is the more reliable route to culturally adequate speech technology in low-resource settings, and they argue for treating cultural adequacy as a standard criterion in low-resource speech evaluation rather than an afterthought.