Assess MBR decoding on real-world noisy speech from diverse communities and regions

Evaluate Minimum Bayes Risk decoding for automatic speech recognition using real-world noisy speech datasets drawn from diverse communities and regions to determine robustness beyond MUSAN-based synthetic noise.

Background

The paper induces noise using the MUSAN corpus to paper robustness, but MUSAN may not capture the full variability and characteristics of real-world noise in different locales and environments.

The authors explicitly note that evaluation on real-world noisy datasets for particular communities and regions is future work, leaving open how MBR performs under authentic noise conditions.