Training Neural Networks with Different Perceptual Loss for Speech Super-Resolution

Authors: Yupeng Shi, Jiatong Shi, Xi Chen, Nengheng Zheng

Audio Samples

Descriptions

WB: the reference wideband speech

NB: the narrowband speech compressed by G.711

GAN_MSE: the estimated wideband speech from the GAN-based SSR system trained by Mean Square Error Loss.

GAN_PWF: the estimated wideband speech from the GAN-based SSR system trained by Perceptual Weighting Filter Loss.

GAN_PM: the estimated wideband speech from the GAN-based SSR system trained by Psychoacoustic Masking Loss.

GAN_PE: the estimated wideband speech from the GAN-based SSR system trained by Perceptual Entropy Loss.

	WB	NB	GAN_MSE	GAN_PWF	GAN_PM	GAN_PE
S1
S2
S3
S4
S5
S6
S7
S8
S9
S10

Spectrogram

Some spectrogram samples are also shown intuitively.

Table 1

Table 2

Table 3

Table 4