New speech recognition benchmark reveals academic tests miss real-world challenges like custom vocabulary that matters most to users

Contextual Earnings-22 dataset created to address gap between academic benchmarks and actual industrial speech-to-text performance

Research shows current academic benchmarks focus on common vocabulary while ignoring rare, context-specific terms that significantly impact transcript usability

Two approaches tested: keyword prompting and keyword boosting both show significant accuracy improvements when properly scaled

Dataset built on Earnings-22 corpus includes realistic custom vocabulary contexts to enable more relevant speech recognition research