New MGhana-ST Dataset Boosts Speech Translation for Ghanaian Languages

Researchers have introduced MGhana-ST, a novel speech translation dataset specifically designed for four low-resource Ghanaian languages: Ga, Twi (Akuapem and Asante dialects), Ewe, and Fante. This dataset represents a significant step towards improving AI's ability to process and translate African languages, which are often underrepresented in global language technology.

The dataset, currently comprising 16.1 hours of paired speech and English translations, was meticulously curated from existing Ghanaian audio resources. A key innovation is that 37 native-speaker annotators directly translated the audio into English, capturing nuanced verbal and non-verbal events, ensuring high-quality, contextually rich data.

Experiments conducted using the Whisper-small model explored the trade-offs between monolingual and multilingual training in data-scarce environments. The findings indicate that flat multilingual training does not universally benefit all varieties; while Ga and Twi showed negligible change, Ewe and Fante experienced performance degradation. This highlights the complexities of cross-lingual transfer, particularly for linguistically distinct or extremely low-resource languages.

Beyond the dataset itself, the research also yielded an important methodological insight: the need for replication across multiple seeds in AI experiments. An earlier single-run analysis had suggested positive transfer for some languages, but this was not sustained across multiple runs, revealing that apparent gains could sometimes be attributed to seed variation rather than genuine transfer. The release of MGhana-ST is poised to catalyze further research in African language speech technology and low-resource speech translation, offering a valuable resource to the global AI community.