Skip to main content
Fig. 4 | EURASIP Journal on Audio, Speech, and Music Processing

Fig. 4

From: Wise teachers train better DNN acoustic models

Fig. 4

Distribution of singular values of hidden layer weights. RIC shown for 6 ×2048 and 5 ×512 networks trained with hard and soft targets. Soft target-trained networks have a slower decay in singular values, requiring more singular values to be retained for a given value of RIC compared to hard target-trained networks

Back to article page