Sound-Classification-Comparison
Classify sound using Deep Learning on Tensorflow and various machine learning models.
Features
- log = Log scale spectrogram
- spectrogram = spectrogram
- features = concat(power spectrum, log-energy, mfcc). Check featuring/features.py
Datasets
Models
Light Gradient Boosting
- log
- spectrogram
eXtreme Gradient Boosting
- log
Bayes
- spectrogram
Deep-learning
- features on feed-forward
- log on convolutional neural network
- log on feed-forward
- spectrogram on convolutional neural network
- spectrogram on feed-forward
- spectrogram on recurrent neural network
Results
Spectrogram Urban Sound
- feed-forward neural network
testing accuracy: 0.463215
precision recall f1-score support
air_conditioner 0.41 0.32 0.36 105
car_horn 0.46 0.40 0.43 58
children_playing 0.33 0.12 0.17 126
dog_bark 0.67 0.58 0.62 134
drilling 0.42 0.72 0.53 137
engine_idling 0.54 0.30 0.39 127
gun_shot 0.60 0.59 0.60 59
jackhammer 0.36 0.73 0.49 122
siren 0.57 0.74 0.64 114
street_music 0.31 0.13 0.19 119
avg / total 0.46 0.46 0.44 1101
- convolutional neural network
testing accuracy: 0.4
precision recall f1-score support
air_conditioner 0.40 0.20 0.27 10
car_horn 0.29 0.33 0.31 6
children_playing 0.16 0.33 0.21 9
dog_bark 0.36 0.31 0.33 16
drilling 0.50 0.55 0.52 11
engine_idling 0.56 0.56 0.56 9
gun_shot 0.33 0.33 0.33 3
jackhammer 0.57 0.44 0.50 9
siren 0.46 0.67 0.55 9
street_music 0.67 0.31 0.42 13
avg / total 0.45 0.40 0.40 95
- recurrent neural network
testing accuracy: 0.453488
precision recall f1-score support
air_conditioner 0.35 0.55 0.43 11
car_horn 0.50 0.40 0.44 5
children_playing 0.20 0.11 0.14 9
dog_bark 1.00 0.64 0.78 11
drilling 0.38 0.45 0.42 11
engine_idling 1.00 0.44 0.62 9
gun_shot 1.00 0.50 0.67 2
jackhammer 0.50 0.60 0.55 10
siren 0.50 0.22 0.31 9
street_music 0.26 0.56 0.36 9
avg / total 0.54 0.45 0.46 86
Spectrogram Google Dataset
- lgb
accuracy validation set: 0.649594438007
precision recall f1-score support
bed 0.56 0.52 0.54 332
bird 0.84 0.66 0.74 381
cat 0.64 0.68 0.66 353
dog 0.61 0.51 0.55 332
down 0.61 0.55 0.58 476
eight 0.68 0.73 0.70 470
five 0.64 0.57 0.60 466
four 0.67 0.77 0.71 469
go 0.52 0.54 0.53 478
happy 0.87 0.84 0.85 368
house 0.78 0.64 0.70 347
left 0.59 0.60 0.59 484
marvin 0.77 0.78 0.78 361
nine 0.63 0.66 0.64 480
no 0.61 0.57 0.59 456
off 0.59 0.67 0.63 456
on 0.67 0.59 0.63 482
one 0.65 0.71 0.68 481
right 0.69 0.62 0.65 446
seven 0.76 0.72 0.74 485
sheila 0.89 0.72 0.80 329
six 0.62 0.81 0.71 490
stop 0.63 0.56 0.59 466
three 0.56 0.56 0.56 446
tree 0.71 0.42 0.53 362
two 0.48 0.70 0.57 478
up 0.57 0.74 0.64 474
wow 0.65 0.59 0.62 343
yes 0.67 0.64 0.65 461
zero 0.75 0.75 0.75 493
avg / total 0.66 0.65 0.65 12945
- bayes
accuracy validation set: 0.245886442642
precision recall f1-score support
bed 0.23 0.31 0.27 359
bird 0.51 0.35 0.42 361
cat 0.23 0.26 0.25 351
dog 0.12 0.13 0.13 348
down 0.36 0.04 0.07 469
eight 0.46 0.34 0.39 464
five 0.12 0.13 0.13 485
four 0.29 0.52 0.38 444
go 0.17 0.11 0.13 474
happy 0.24 0.22 0.23 339
house 0.26 0.40 0.31 332
left 0.11 0.28 0.16 474
marvin 0.19 0.19 0.19 344
nine 0.32 0.22 0.26 450
no 0.20 0.09 0.13 484
off 0.26 0.42 0.32 468
on 0.38 0.10 0.16 473
one 0.18 0.15 0.16 481
right 0.37 0.05 0.09 472
seven 0.52 0.19 0.28 457
sheila 0.34 0.39 0.36 354
six 0.37 0.47 0.42 476
stop 0.30 0.27 0.28 507
three 0.27 0.30 0.28 481
tree 0.31 0.38 0.34 347
two 0.28 0.31 0.30 543
up 0.32 0.07 0.12 490
wow 0.11 0.20 0.14 314
yes 0.16 0.26 0.20 413
zero 0.23 0.30 0.26 491
avg / total 0.28 0.25 0.24 12945
Log-spectrogram Google Dataset
- LGB
accuracy validation set: 0.722689742775
precision recall f1-score support
bed 0.64 0.68 0.66 359
bird 0.82 0.74 0.78 327
cat 0.77 0.72 0.74 353
dog 0.78 0.59 0.67 340
down 0.67 0.67 0.67 457
eight 0.75 0.79 0.77 466
five 0.73 0.65 0.69 514
four 0.78 0.82 0.80 477
go 0.54 0.61 0.57 463
house 0.87 0.75 0.80 324
left 0.75 0.70 0.72 488
marvin 0.79 0.76 0.78 370
nine 0.70 0.75 0.72 447
no 0.70 0.57 0.63 518
off 0.71 0.74 0.72 450
on 0.66 0.66 0.66 466
one 0.72 0.75 0.73 485
right 0.80 0.73 0.76 453
seven 0.78 0.73 0.75 467
sheila 0.88 0.81 0.85 362
six 0.69 0.85 0.76 469
stop 0.77 0.73 0.75 452
three 0.68 0.71 0.70 461
tree 0.80 0.62 0.70 341
two 0.65 0.81 0.72 530
up 0.59 0.80 0.68 470
wow 0.80 0.71 0.76 358
yes 0.75 0.70 0.72 470
zero 0.76 0.78 0.77 459
avg / total 0.73 0.72 0.72 12596
- XGB
accuracy validation set: 0.801047951731
precision recall f1-score support
bed 0.71 0.78 0.74 330
bird 0.85 0.83 0.84 348
cat 0.81 0.85 0.83 362
dog 0.76 0.70 0.73 351
down 0.77 0.74 0.75 442
eight 0.83 0.86 0.85 485
five 0.80 0.77 0.78 478
four 0.83 0.88 0.86 460
go 0.68 0.69 0.69 460
house 0.90 0.82 0.86 368
left 0.81 0.78 0.79 495
marvin 0.87 0.80 0.84 363
nine 0.81 0.80 0.80 455
no 0.76 0.74 0.75 468
off 0.77 0.78 0.78 464
on 0.78 0.74 0.76 447
one 0.76 0.82 0.79 418
right 0.86 0.80 0.83 491
seven 0.86 0.82 0.84 495
sheila 0.94 0.87 0.90 359
six 0.78 0.88 0.83 474
stop 0.84 0.80 0.82 489
three 0.78 0.80 0.79 476
tree 0.81 0.71 0.76 329
two 0.77 0.81 0.79 506
up 0.70 0.82 0.76 504
wow 0.80 0.82 0.81 330
yes 0.84 0.86 0.85 496
zero 0.82 0.84 0.83 453
avg / total 0.80 0.80 0.80 12596
images
