Skip to main content

Table 6 EER (%) for 2 ×1024 network with soft targets from CNN

From: Wise teachers train better DNN acoustic models

Network Targets Data Mob-1 Mob-2 Mob-3 Mob-4
CNN Hard alignment 320 h labeled 4.10 % 2.51 % 2.46 % 2.77 %
MLP baseline Hard alignment 320 h labeled 4.46 % 2.75 % 2.79 % 3.09 %
MLP student CNN outputs 320 h unlabeled 4.26 % 2.61 % 2.60 % 2.91 %
MLP student CNN outputs 640 h unlabeled 4.19 % 2.55 % 2.52 % 2.85 %
MLP student CNN outputs 960 h unlabeled 4.14 % 2.52 % 2.47 % 2.79 %