notesum.ai

Published at October 20

Improving Voice Quality in Speech Anonymization With Just Perception-Informed Losses

cs.AI

cs.SD

eess.AS

Released Date: October 20, 2024

Authors: Suhita Ghosh¹, Tim Thiele¹, Frederic Lorbeer¹, Frank Dreyer¹, Sebastian Stober¹

Aff.: ¹Otto-von-Guericke-University, Magdeburg, Germany

Arxiv: https://arxiv.org/abs/2410.15499v1

Source

All conversions (All)

/Accent-wise

CER [%]

\downarrow

EER [%]

\uparrow

M_{\text{base}}

M_{\text{PL}}

M_{\text{base}}

M_{\text{PL}}

All Conversions

All

71.33

\pm

0.40

53.09

\pm

0.67

41.07

43.21

VCTK

\rightarrow

VCTK

All

73.32

\pm

0.79

45.49

\pm

1.13

37.88

38.17

American

\rightarrow

British

73. 31

\pm

1.49

45.32

\pm

2.24

36.89

39.10

Canadian

\rightarrow

British

72.02

\pm

1.39

48.89

\pm

1.97

37.10

42.23

British

\rightarrow

British

74.30

\pm

1.26

43.07

\pm

1.68

36.07

38.02

LibriSpeech

\rightarrow

VCTK

All

70.16

\pm

0.64

53.50

\pm

1.12

38.26

38.00

German

\rightarrow

VCTK

All

77.44

\pm

0.56

75.21

\pm

0.75

42.35

44.84

German

\rightarrow

German

All

64.41

\pm

0.91

38.15

\pm

1.02

49.63

51.46