Building an Accurate Open-Source Hebrew ASR System through Crowdsourcing

Automatic Speech Recognition (ASR) for Hebrew faces significant challenges due to limited resources and rich morphology. While recent advances have improved high-resource languages ASR, Hebrew still lacks robust open-source solutions. Through crowdsourcing efforts, we created a dataset of 314 hours of transcribed speech, which we used to train a new Hebrew ASR model based on the Whisper architecture. Our model demonstrates up to 29% reduction in error rates compared to existing Whisper solutions, particularly excelling in producing verbatim transcriptions. Additionally, we introduce a new evaluation dataset designed specifically for Hebrew ASR assessment. By making both the model and methodology freely available, we provide a framework that can be adapted for developing ASR systems in other under-resourced languages. This work represents a step toward making speech technology more accessible in different languages.