Table 12.

Experimental results of various post-processing models and the proposed CREAM framework

MethodHigh-error scenarioLow-error scenario
AISHELL-1TEDLIUM-2Libri-otherLibri-cleanAISHELL-1TEDLIUM-2Libri-otherLibri-clean
N-best reranking method
CLM6.8010.619.993.646.558.927.052.84
MLM (BERT)6.0510.379.913.585.558.446.862.73
RescoreBERT6.1210.619.983.695.568.776.962.79
PBERT5.8910.359.973.635.358.516.812.74
LLM-Reranking (ChatGPT 3.5)7.3611.6711.334.908.6011.159.725.72
Error correction model
BART7.4311.7711.264.507.5310.667.423.05
FastCorrect6.726.81
LLM-Correction (ChatGPT 3.5)7.7113.9512.776.077.4014.0210.706.01
CREAM
CREAM (BART)4.639.298.803.164.988.136.722.70
CREAM (FastCorrect)4.734.73

or Create an Account

Close subscription notice
Close access options