The Anatomy of the CTC Oracle Gap: the near-converged CTC ASR decoding bottleneck is linguistic, not acoustic. MBR + RoBERTa PLL → ~9% relative WER reduction. Paper + code
pytorch speech-recognition beam-search asr k2 ctc librispeech roberta mbr-decoding language-model-rescoring
-
Updated
Jun 24, 2026 - Python