EURASIP Journal on Audio, Speech, and Music Processing

Table 8 Rank of hidden layer weight matrices for networks trained with soft vs. hard targets using 110 h of data

From: Wise teachers train better DNN acoustic models

Hidden layer	6 ×2048		5 ×512
	Soft	Hard	Soft	Hard
1	2043	943	511	511
2	2040	731	511	510
3	2040	562	512	499
4	2039	483	512	498
5	2042	487	–	–

Back to article page