Skip to main content

Table 3 UERs for different lexicons: canonical, 2–5 variants with and without priors. These rates are obtained by using non-native triphone acoustic models and language models with filled pause loops.

From: Optimizing Automatic Speech Recognition for Low-Proficient Non-Native Speakers

Lex Priors 0 5 10 15
canonical 10.0 9.7 7.9 6.9
2 var No 10.0 9.9 8.2 6.7
2 var Yes 10.0 9.7 8.3 7.0
3 var No 11.2 10.9 8.5 7.1
3 var Yes 10.6 10.1 8.7 7.2
4 var No 11.5 11.3 8.9 7.5
4 var Yes 10.4 10.9 9.7 7.2
5 var No 11.5 11.3 8.9 7.5
5 var Yes 10.4 10.0 8.7 7.2