Evaluate MBR decoding on non-autoregressive ASR architectures (e.g., CTC-based models)

Evaluate the effectiveness of Minimum Bayes Risk decoding when applied to Connectionist Temporal Classification (CTC)-based automatic speech recognition models and other non-autoregressive probabilistic architectures.

Background

All experiments in the paper use sequence-to-sequence autoregressive models (e.g., Whisper). The authors note that MBR is a general decoding framework that could be applied to other probabilistic models, including CTC-based ASR.

They explicitly state that evaluating MBR on other model types remains future work, leaving open whether similar accuracy gains extend beyond autoregressive systems.