Skip to main content

Table 5 Parameter reduction by soft target training

From: Wise teachers train better DNN acoustic models

Network # of Params Data Hub5’00-SWB RT03S-FSH
6 ×2048 (hard align., sMBR) 30.4 mil. 110 h transcribed (SWBD) 18.3 % 22.6 %
5 ×512 (soft target-trained) 3.4 mil. 1100 h untrans. (SWBD + FSH) 18.0% 21.4%