Skip to main content

Table 4 Performance(TER%) comparison of the proposed lexicon learning-based pronunciation augmentation on conventional ASR systems

From: Pronunciation augmentation for Mandarin-English code-switching speech recognition

System

ρ

Lexicon(avg.#prons)

Mandarin

Code-switching

Chilish

GMM-HMM

Source Lexicon

17.5

23.7

27.0

 

0.1

+ LexLearn (3.65)

17.8

22.8

26.4

 

0.4

+ LexLearn (3.07)

17.6

22.1

26.2

 

0.7

+ LexLearn (2.41)

17.6

21.6

26.0

LF-MMI hybrid

Source Lexicon

5.2

11.5

15.4

LF-MMI hybrid-1

0.7

+ LexLearn (2.41)

5.7

10.9

15.2

LF-MMI hybrid-2

0.7

+ LexLearn (2.41)

5.5

10.4

14.8

  1. “avg.#prons” means the average pronunciations per English word included in the training data. ρ is the pruning factor of acoustic soft counts in Eq.(2)