Pioneering AI Research Develops First Text-to-Speech System for Nigeria's Efik Language

A new research study marks a significant step towards the digital preservation of Efik, a tonal language primarily spoken in Southeastern Nigeria. With approximately 1.5 million native speakers and 3 million second-language speakers, Efik has historically been underrepresented in speech synthesis research, limiting its digital accessibility and potential for integration into modern AI applications.

The researchers conducted the first documented end-to-end text-to-speech (TTS) study for Efik. This involved creating a new, curated corpus of 2,632 utterances, totaling three hours of speech from a single speaker. They then compared the performance of four neural TTS models—VITS, MMS-TTS, SpeechT5, and Orpheus-TTS—under low-resource conditions, assessing their ability to accurately synthesize Efik speech.

Native Efik speakers were engaged to evaluate the systems using standard metrics like Mean Opinion Score (MOS). The MMS-TTS model demonstrated the highest MOS of 3.80 and proved more stable for generating longer speech segments. However, the study highlighted persistent challenges, particularly with tonal accuracy across all models, indicating the complex nature of synthesizing tonal languages.

This work provides a crucial reproducible baseline for future research into Efik TTS and other low-resource tonal African languages. The findings underscore the urgent need for larger, more diverse speech corpora and the development of AI models specifically designed to handle the intricate tonal and prosodic features characteristic of many African languages.

For Africa, this research is vital as it contributes directly to the digital inclusion and cultural preservation of indigenous languages. Developing robust AI tools for languages like Efik can unlock new possibilities for education, communication, and economic development within local communities, bridging the digital divide for millions.