EURASIP Journal on Audio, Speech, and Music Processing

Table 1 Reconstruction SDR under different positive constraints in decibel

From: Discriminative frequency filter banks learning with neural networks

Initialization	Constraint	M = 32		M = 64
		F-B	T-I	F-B	T-I
N(0, 0.1)	Exponent	8.57	6.72	13.10	6.74
	Sigmoid	8.54	8.89	12.84	9.43
	ReLU	8.45	14.44	12.84	18.54
	Square	8.57	14.44	12.84	18.24
N(− 3.0, 2.0)	Exponent	9.02	14.44	13.25	17.70
	Sigmoid	8.36	14.31	12.84	17.97

F-B represents the parameters with fixed frequency center and bandwidth; T-I represents the totally independent parameters; M represents the number of filters; N means Gaussian distribution

Back to article page