A new research paper covered by arXiv cs.AI details an end-to-end study evaluating automatic speech recognition for adolescent health communication across three Ghanaian languages: Twi, Dagbani, and Ewe. The investigation involves testing five distinct automatic speech recognition systems on both a general-domain Bible corpus and a specialized Youth Adolescent Sexual and Reproductive Health domain dataset, measuring performance through Character Error Rate and Word Error Rate metrics.
Evaluating Multimodal LLMs and Language-Specific Models
The study examines five different automatic speech recognition systems, comprising three language-specific Wav2Vec2 models alongside two multimodal large language models identified as Gemma 3n and Gemma 4. Following the initial benchmarking phase, the research applies supervised domain adaptation guided by the performance data gathered during the evaluations.
Unanswered Questions in the Research
The available information does not specify the complete architectural details of the supervised domain adaptation methods applied after the benchmarking phase. Furthermore, the exact performance scores for each of the five tested systems, the specific size of the datasets used, and the full scope of Gemma 4’s capabilities prior to adaptation were not detailed in the report.
Further updates on this research will depend on the full release of the paper and its accompanying datasets.
Lead image: Grish Petrosyan / Pexels
