DCASE 2016 CONVOLUTIONAL NEURAL NETWORKS FOR ACOUSTIC SCENE - - PowerPoint PPT Presentation
DCASE 2016 CONVOLUTIONAL NEURAL NETWORKS FOR ACOUSTIC SCENE CLASSIFICATION Michele Valenti 1 (valenti.michele.w@gmail.com), Aleksandr Diment 2 , Giambattista Parascandolo 2 , Stefano Squartini 1 , Tuomas Virtanen 2 1 Universit Politecnica delle
Michele Valenti1 (valenti.michele.w@gmail.com), Aleksandr Diment2, Giambattista Parascandolo2, Stefano Squartini1, Tuomas Virtanen2
1Università Politecnica delle Marche, Italy 2Tampere University of T
echnology, Finland
Michele Valenti1 (valenti.michele.w@gmail.com), Aleksandr Diment2, Giambattista Parascandolo2, Stefano Squartini1, Tuomas Virtanen2
1Università Politecnica delle Marche, Italy 2Tampere University of T
echnology, Finland
What is “acoustic scene classification”?
Forest path Car Home Audio
What is “acoustic scene classification”?
Audio Feature extraction Sequence splitting Label CNN Scores averaging Overview
Log-mel spectrogram Raw audio Features
Audio Features
Log-mel spectrogram Raw audio segment Sequence Sequence splitting
Sequence splitting Features
Sequence Convolutional neural network
Sequence Feature maps 128 Convolutional neural network
CNN Sequences
Sequence Feature maps 128 Convolutional neural network
CNN Sequences
Batch normalization
Sequence 128 128 Feature maps Subsampled feature maps Convolutional neural network
CNN Sequences
Feature maps Sequence Subsampled feature maps 128 128 256 New feature maps Convolutional neural network
CNN Sequences
Sequence 256 128 128 Feature maps Subsampled feature maps New feature maps
Time shrinking
Convolutional neural network
CNN Sequences
Sequence 256 128 128
Flattening
Feature maps Subsampled feature maps New feature maps Convolutional neural network
CNN Sequences
Sequence 256 128 128
Fully-connected softmax layer
Feature maps Subsampled feature maps New feature maps Convolutional neural network
CNN Sequences
Sequence 256 128 128 Feature maps Subsampled feature maps New feature maps Convolutional neural network
CNN Sequences
Scores averaging
Class prediction scores
Scores averaging Prediction scores
Scores averaging
Class prediction scores File’s class
argmax
Scores averaging Prediction scores
Cross-validation setup
T est Training + validation T est T est T est Fold 1 Fold 2 Fold 3 Fold 4
Non-full training Training Validation
T raining + validation T est Fold n
Non-full training Training Validation
T raining + validation T est Fold n
Non-full training
Non-full training Training Validation Accuracies Epochs Training
T raining + validation T est Fold n
Validation
Non-full training Training Validation Accuracies Epochs Training
T raining + validation T est Fold n
Validation
Convergence time
Non-full training Training Validation
T raining + validation T est Fold n
Training
Non-full training Training Validation
T raining + validation T est Fold n
Training
Full training
T est Training + validation T est T est T est Fold 1 Fold 2 Fold 3 Fold 4
Test data
Sequence length
65 70 75 80 0,5 1,5 3 5 10 30
Accuracy (%) Sequence length (s) Non-full training Full training
Sequence length
65 70 75 80 0,5 1,5 3 5 10 30
Accuracy (%) Sequence length (s) Non-full training Full training
Sequence length
65 70 75 80 0,5 1,5 3 5 10 30
Accuracy (%) Sequence length (s) Non-full training Full training
Class accuracies
Class Accuracy (%) Beach 75.6 Bus 76.9 Café/Restaurant 74.4 Car 91.0 City center 93.6 Forest path 96.2 Grocery store 88.5 Home 80.8 Class Accuracy (%) Library 66.6 Metro station 96.2 Office 97.4 Park 59.0 Residential area 73.1 T rain 46.2 T ram 78.2
Class Accuracy (%) Beach 75.6 Bus 76.9 Café/Restaurant 74.4 Car 91.0 City center 93.6 Forest path 96.2 Grocery store 88.5 Home 80.8 Class Accuracy (%) Library 66.6 Metro station 96.2 Office 97.4 Park 59.0 Residential area 73.1 Train 46.2 Tram 78.2
Class accuracies
34.6% Residential area 29.5% Bus
Other classifiers
System Sequence length (s) Accuracy (%) Non-full training Full training Baseline GMM (MFCC)
T wo-layer CNN (MFCC) 5 67.7 72.6 T wo-layer MLP (log-mel)
69.3 One-layer CNN (log-mel) 3 70.3 74.8 Two-layer CNN (log-mel) 3 75.9 79.0
Final training
Training + validation + test Secret challenge data Extended training set Evaluation set
Final training
Training + validation + test Extended training set Evaluation set New training New validation Secret challenge data
Final training
Training + validation + test Extended training set Evaluation set New training New validation Secret challenge data
400 epochs convergence
Final training
Training + validation + test Extended training set Evaluation set Final training for 400 epochs Secret challenge data
89,7 88,7 87,7 87,2 86,4 86,4 86,2 85,9 85,6 85,4 84,6 84,1 77,2 62,8 10 20 30 40 50 60 70 80 90 100
Michele Valenti1 (valenti.michele.w@gmail.com), Aleksandr Diment2, Giambattista Parascandolo2, Stefano Squartini1, Tuomas Virtanen2
1Università Politecnica delle Marche, Italy 2Tampere University of T
echnology, Finland
Feature comparison
System Sequence length (s) Accuracy (%) Non-full training Full training T wo-layer CNN (MFCC) 5 67.7 72.6 T wo-layer CNN (log-mel) 5 74.1 78.3