notesum.ai

Published at November 11

Wavehax: Aliasing-Free Neural Waveform Synthesis Based on 2D Convolution and Harmonic Prior for Reliable Complex Spectrogram Estimation

cs.SD

eess.AS

Released Date: November 11, 2024

Authors: Reo Yoneyama¹, Atsushi Miyashita¹, Ryuichi Yamamoto¹, Tomoki Toda¹

Aff.: ¹Not specified

Arxiv: http://arxiv.org/abs/2411.06807v1

	Prior	Input type	Output type	Subset	VUV $\downarrow$	RMSE $\downarrow$	STFT $\downarrow$	PESQ $\uparrow$	UTMOS $\uparrow$
N-RI-RI	Noise	Real Imaginary	Real Imaginary	I	10	0.111	0.711	3.108	2.038
				II	34	0.212	0.821	2.288	1.465
				III	4	0.078	0.763	3.306	1.424
S-RI-RI	Sine	Real Imaginary	Real Imaginary	I	6	0.085	0.701	3.733	2.972
				II	13	0.127	0.708	3.589	3.326
				III	4	0.072	0.839	3.209	1.524
H-RI-RI	Harmonic	Real Imaginary	Real Imaginary	I	6	0.081	0.678	3.818	3.122
				II	13	0.114	0.683	3.702	3.585
				III	4	0.071	0.722	3.927	1.686
H-LA-LP	Harmonic	Log amplitude Absolute phase	Log amplitude IPW	I	6	0.080	0.682	3.793	3.071
				II	13	0.117	0.687	3.684	3.531
				III	4	0.070	0.714	3.934	1.686
H-L-LP	Harmonic	Log amplitude	Log amplitude IPW	I	8	0.098	0.683	3.47	2.309
				II	22	0.167	0.712	2.84	1.994
				III	4	0.075	0.720	3.67	1.524