notesum.ai

Published at December 4

PERL: Pinyin Enhanced Rephrasing Language Model for Chinese ASR N-best Error Correction

cs.CL

Released Date: December 4, 2024

Authors: Junhong Liang¹

Aff.: ¹State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, CAS, Beijing, China

Arxiv: http://arxiv.org/pdf/2412.03230v1

Method

CER%

\downarrow

_-CERR%↓

Aishell1

Law

Med

Odw

Baseline (1-best)

5.84

12.64

16.09

15.94

O_{cp}

2.93 _-49.83

3.83 _-69.70

6.13 _-61.90

4.41 _-73.59

O_{nb}

3.49 _-32.70

9.44 _-25.31

12.66 _-28.37

11.89 _-28.80

ReLM

5.20_-10.96

5.21_-58.78

6.91_-57.05

6.17_-61.23

GPT-4o

4.41_-24.49

9.83_-22.23

13.27_-17.53

12.40_-22.20

DeepSeek-V2.5

4.25_-27.22

9.12_-27.85

12.61_-21.63

11.68_-26.72

Qwen2.5

4.22_-27.74

8.28_-34.49

12.16_-24.43

11.15_-30.05

PERL

4.10_-29.11

3.41_-73.02

3.91_-75.70

4.87_-69.44