Evaluation#

We train a convnet over 3 acoustic feature representations: the CQT, time scattering coefficients and joint time-frequency scattering, and evaluate the classification accuracy and confusions between playing techniques.

CQT#

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Test metric ┃ DataLoader 0 ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ test/acc │ 0.723718523979187 │ │ test/loss │ 0.6989585161209106 │ └───────────────────────────┴───────────────────────────┘

Confusion Matrix on test set for CQT model

Scat1D#

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Test metric ┃ DataLoader 0 ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ test/acc │ 0.9504778385162354 │ │ test/loss │ 0.14028948545455933 │ └───────────────────────────┴───────────────────────────┘

Confusion Matrix on test set for Scattering1D model

JTFS#

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Test metric ┃ DataLoader 0 ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ test/acc │ 0.9704604744911194 │ │ test/loss │ 0.07301218807697296 │ └───────────────────────────┴───────────────────────────┘

Confusion Matrix on test set for JTFS model