Skip to main content
. 2022 May 12;22(10):3683. doi: 10.3390/s22103683

Table 6.

The E2E − T Transformer and LM rescoring setups were compared (WER % and RT factor). Transcription (TR) and background (BG) texts were used to train the language models.

Methods Beam 20 Beam 5 Beam 2
WER RT WER RT WER RT
E2E − T + LSTM − LM (BG + TR) 15.4 4.46 16.9 1.89 16.6 1.87
E2E − T + T − LM (BG + TR) 14.7 6.37 15.7 2.62 15.9 1.96
E2E − T + T − LM (TR) 14.6 6.36 15.4 2.53 15.5 1.71
E2E − T 14.3 4.31 15.2 1.97 15.3 1.48