The real-world usability of automatic speech recognition (ASR) models remains a challenge, particularly for low-resource languages. Existing speech resources for such languages are often collected in clean studio environments, which do not reflect the conditions of actual telecommunication, such as background noise, compression, and varying audio quality.
To address this gap, the paper introduces TeleASR-Amharic, a benchmark designed for robust Amharic ASR under realistic telecommunication conditions. The benchmark aims to provide a more practical evaluation setting for Amharic, a low-resource language, moving beyond the controlled studio recordings that dominate current datasets.
The significance of this work lies in bridging the gap between laboratory performance and real-world deployment. By focusing on telecommunication conditions, the benchmark could help develop ASR systems that work better in everyday phone calls and similar scenarios, ultimately improving accessibility for Amharic speakers.