Table 1 Average performance of each model according to accuracy, specificity at 90% sensitivity, the ROC AUC, and ROC AUC mean difference from baseline model

Model 0

Baseline

88.7% [87.8, 89.6]

90.1% [89.7, 90.5]

85.1% [82.7, 87.5]

0.943 [0.935, 0.956]

–

Model 1

Pitch/time alterations

88.7% [87.6, 89.8]

–

90.2% [89.8, 90.6] ↑

81.3% [75.3, 87.3] ↓

0.925 [0.918, 0.935] ↓

− 0.018 [− 0.031, − 0.004]

Model 2

Noise injection

88.6% [87.7, 89.5] ↓

90.1% [89.9, 90.3]

–

82.1% [77.3, 86.9] ↓

0.932 [0.915, 0.943] ↓

− 0.011 [− 0.023, 0.001]

Model 3.1

Horizontal flip

90.7% [89.8, 91.6] ↑

89.5% [89.2, 89.8] ↓

90.4% [88.9, 91.9] ↑

0.956 [0.952, 0.964] ↑

0.013 [0.007, 0.018]

Model 3.2

Vertical flip

88.9% [87.5, 90.3] ↑

90.1% [89.8, 90.4]

–

72.8% [64.7, 80.9] ↓

0.920 [0.908, 0.930] ↓

− 0.023 [− 0.037, − 0.008]

Model 4.1

SV perturbations

90.7% [89.6, 91.8] ↑

90.0% [89.6, 90.4] ↓

77.5% [61.3, 93.7] ↓

0.940 [0.933, 0.958] ↑

− 0.004 [− 0.017,0.009]

Model 4.2

PCA color augmentation

89.3% [88.5, 90.1] ↑

90.3% [90.0, 90.6] ↑

87.6% [84.9, 90.3] ↑

0.941 [0.941, 0.958] ↓

− 0.002 [− 0.013,0.008]

Model 4.3

Random color filters

89.1% [87.7, 90.5] ↑

90.0% [89.6, 90.4] ↓

85.1% [80.9, 89.3]

–

0.938 [0.912, 0.943] ↓

− 0.006 [− 0.018,0.007]

Model 5

Time/frequency masking

88.7% [87.6, 89.8]

–

90.1% [89.7, 90.5]

–

83.0% [79.6, 86.4] ↓

0.934 [0.941, 0.956] ↓

− 0.010 [− 0.020,0.002]

Model 6

Horizontal flip and PCA

91.0% [90.0, 92.0] ↑

89.9% [89.7, 90.1] ↓

90.8% [88.8, 92.8] ↑

0.958 [0.949, 0.968] ↑

0.015 [0.006,0.023]

Model 7

Horizontal flip and SV perturbations

90.7% [89.8, 91.6] ↑

90.2% [89.8, 90.6] ↑

91.0% [90.0, 92.0] ↑

0.955 [0.948, 0.962] ↑

0.012 [0.004,0.019]