New AI Text-to-Speech Benchmark Prioritizes Underrepresented Languages, Showing Strong Performance for African Tongues

Recent advancements in neural text-to-speech (TTS) technology have dramatically improved synthetic speech quality, yet these benefits have largely bypassed low-resource languages, which often lack the extensive data needed for training robust AI models. Existing TTS systems predominantly cater to a small number of high-resource languages, leaving a significant gap for linguistic communities with fewer digital resources. To address this disparity, researchers have introduced OpenBibleTTS, a comprehensive, large-scale benchmark designed specifically for low-resource speech synthesis.

OpenBibleTTS encompasses 37 underrepresented languages, providing a crucial dataset for developing more inclusive AI speech technologies. The project involved a systematic evaluation of various TTS architectures and large-scale speech generation models, testing their performance on both in-domain (Biblical) and out-of-domain text. This rigorous comparison highlights the complexities of achieving consistent quality across diverse linguistic contexts, demonstrating that no single AI system universally outperforms others.

Significantly, the study found that while Gemini-TTS excelled in listener ratings for many languages, monolingual EveryVoice models, specifically trained on the OpenBibleTTS dataset, proved superior in terms of intelligibility and were preferred by users in several African languages. This finding underscores the potential for tailored AI solutions to meet the unique needs of specific linguistic groups. However, the research also revealed a persistent challenge: open-source systems trained from scratch often struggle with out-of-domain text, indicating a gap between broad multilingual coverage and reliable synthesis quality in underserved communities.

For Africa, this research is particularly impactful as it directly addresses the digital language divide that often excludes African languages from mainstream AI applications. By providing open-source datasets, alignments, and trained models, OpenBibleTTS empowers researchers and developers to build more effective and culturally relevant speech technologies for African populations. The strong performance of certain models in African languages signals a promising step towards more inclusive AI that can support local communication, education, and digital access across the continent.