Urban sound source tagging from an aggregation of four second noisy audio clips via 1D and 2D CNN (Xception)
Urban sound source tagging from an aggregation of four second noisy audio clips via 1D and 2D CNN (Xception)
The Urban Sound Classification dataset contains 8732 labeled sound excerpts (<=4s) of urban sounds from 10 classes,namely:
The attributes of data are mapped as follows:
.
├── data
│ ├── img
│ │ ├── audio-features.png
│ │ ├── sound.png
│ │ └── time_freq.png
│ ├── test
│ │ └── Test
| | ├── 1.wav
| | ├── 2.wav
| | ├── .............
│ ├── test.csv
│ ├── train
│ │ └── Train
| | ├── 1.wav
| | ├── 2.wav
| | ├── ............
| |
│ └── train.csv
├── LICENSE
├── notebooks
│ ├── eda_plots
│ │ ├── amplitude_vs_time
│ │ │ ├── air_conditioner.svg
│ │ │ ├── car_horn.svg
| | | ├── ............
│ │ └── mel_spectrum
│ │ ├── air_conditioner.png
│ │ ├── car_horn.png
| | ├── ............
│ └── Exploratory Data Analysis.ipynb
├── README.md
├── requirements.txt
├── results
│ ├── acc_model_1d.png
│ ├── acc_model_2d.png
│ ├── loss_model_1d.png
│ ├── loss_model_2d.png
│ ├── pred_1d.csv
│ └── pred_2d.csv
└── src
├── test_1d.py
├── test_2d.py
├── train_1d.py
├── train_2d.py
├── utils_1d.py
└── utils_2d.py
Exploratory Data Analysis:
Audio Tagging:
Normalizing the audio clips and passing them through stacks of 1D convolution layers for feature extraction. Then the usual dense layer stacks were used to do the final categorization.
Extracting features in the form of mel-spectogram and passing them through stacks of 2D convolution layers for additional feature pulling. Dense layer stack does the final classification. In this case, we trained an Xception model from scratch to achieve better generalization capability.
We achieved 89% validation accuracy in the second approach.
pip install -r requirements.txt