Skip to main content

Table 3 Performance of the proposed and baseline systems

From: Music detection from broadcast contents using convolutional neural networks with a Mel-scale kernel

Test data

Model type

F-score (%)

Precision (%)

Recall (%)

Korean drama (dev)

Spectrogram + CNN with melCL (proposed)

95.9

95.9

96.0

Spectrogram + CNN

92.2

94.0

90.5

Mel-spectrogram + CNN

94.2

95.7

92.8

Spectrogram + bi-GRU

88.0

87.0

89.0

Mel-spectrogram + bi-GRU

93.4

91.9

95.0

Mel-spectrogram + bi-LSTM

90.6

90.1

91.1

Korean reality

Spectrogram + CNN with melCL (proposed)

94.7

93.0

96.4

Spectrogram + CNN

90.7

91.4

89.9

Mel-spectrogram + CNN

93.5

91.1

95.9

spectrogram + bi-GRU

90.6

84.9

97.2

Mel-spectrogram + bi-GRU

92.3

88.5

87.8

Mel-spectrogram + bi-LSTM

92.6

87.5

98.4

British 8 h

Spectrogram + CNN with melCL (proposed)

86.5

85.3

87.8

Spectrogram + CNN

83.5

79.8

87.5

Mel-spectrogram + CNN

86.8

83.3

90.5

Spectrogram + bi-GRU

75.0

65.7

87.4

Mel-spectrogram + bi-GRU

78.5

67.8

93.1

Mel-spectrogram + bi-LSTM

80.5

72.5

90.5

Spanish 12 h

Spectrogram + CNN with melCL (proposed)

88.9

84.7

93.4

Spectrogram + CNN

86.6

80.0

94.4

Mel-spectrogram + CNN

80.9

70.6

94.6

Spectrogram + bi-GRU

75.3

63.8

92.0

Mel-spectrogram + bi-GRU

74.1

61.5

93.2

Mel-spectrogram + bi-LSTM

75.6

63.4

93.6

MIREX

2015

Spectrogram + CNN with melCL (proposed)

95.3

99.4

91.6

Spectrogram + CNN

93.8

98.8

89.3

Mel-spectrogram + CNN

92.5

93.8

91.2

Spectrogram + bi-GRU

92.8

94.9

90.8

Mel-spectrogram + bi-GRU

94.3

92.3

96.4

Mel-spectrogram + bi-LSTM

95.3

94.1

92.7

Dafx 07

Spectrogram + CNN with melCL (proposed)

84.9

84.0

85.9

Spectrogram + CNN

84.4

77.7

92.3

Mel-spectrogram + CNN

80.1

69.2

95.1

Spectrogram + bi-GRU

68.4

57.5

84.5

Mel-spectrogram + bi-GRU

69.0

53.3

98.0

Mel-spectrogram + bi-LSTM

70.6

55.4

97.3