Figure 8
From: A perceptual masking approach for noise robust speech recognition
Comparison of average WER (%) performance calculated from 0 to 20 dB for spectral subtraction, ETSI-2 frontend and proposed approach for test sets A, B, and C. The first and fourth and second and fifth bars indicate performance of spectral subtraction and ETSI-2 frontend for training on clean data and multi-condition data. The third and sixth indicate performance of proposed approach for training on clean data and multi-condition data.