Evaluating the impact of different acoustic contexts on German speech recognition

Recent advances in Automatic Speech Recognition (ASR) research suggest that the error rates have gradually decreased, even for low-resourced languages. However, the performance of ASR methods can vary significantly depending on the quality of the audio signal and the presence of noise or other distortions. The signal degradation and the added noise are a direct consequence of the acoustic context in which the audio was recorded. While the robustness of ASR methods w.r.t. noise is very relevant for practical applications, it is rarely evaluated systematically under diverse environmental and acoustical contexts.In this paper, we propose an evaluation strategy with specific environmental contexts for ASR methods that involve adding relevant ambient noise and room reverberations to a clean dataset and measuring the impact of acoustic contexts on recognition accuracy. Our method allows for a more comprehensive and systematic assessment of the robustness of ASR models under specific real-world contexts. We use our evaluation method to compare a number of German ASR models and show that it can provide valuable insights into the performance of ASR models under different contexts This also compels the research community to explore more methods for domain adaptation.