As the final project of Global AI Hub DL Bootcamp, we used Convolutional Neural Network (CNN) to classify urban sounds.
The project includes the following steps:
- Creating Mel spectrograms from sound files
- Preprocessing the spectrograms for training (grayscaling - resizing - normalization)
- Splitting the data into train, test, and validation sets
- Model building and training
The dataset contains 8732 labeled sound excerpts of urban sounds from 10 classes and it is publicly available here.
A random sample from the metadata is given below:
| slice_file_name | fsID | start | end | salience | fold | classID | class | |
|---|---|---|---|---|---|---|---|---|
| 892 | 118278-4-0-7.wav | 118278 | 3.5 | 7.5 | 2 | 10 | 4 | drilling |
| 6352 | 39967-9-0-0.wav | 39967 | 19.3035 | 23.3035 | 1 | 8 | 9 | street_music |
| 1024 | 123688-8-0-0.wav | 123688 | 14.1756 | 18.1756 | 2 | 2 | 8 | siren |
| 2625 | 159738-8-0-12.wav | 159738 | 6.74699 | 10.747 | 2 | 1 | 8 | siren |
| 1581 | 138465-1-0-0.wav | 138465 | 19.2123 | 22.9625 | 2 | 8 | 1 | car_horn |
| 5766 | 24347-8-0-20.wav | 24347 | 14.1204 | 18.1204 | 2 | 4 | 8 | siren |
| 7888 | 75743-0-0-22.wav | 75743 | 11 | 15 | 2 | 9 | 0 | air_conditioner |
| 8293 | 85249-2-0-68.wav | 85249 | 34 | 38 | 1 | 6 | 2 | children_playing |
| 1347 | 13230-0-0-7.wav | 13230 | 3.5 | 7.5 | 1 | 3 | 0 | air_conditioner |
| 2734 | 159751-8-0-11.wav | 159751 | 6.6505 | 10.6505 | 2 | 4 | 8 | siren |
A spectrogram is a visualization of the frequency spectrum of a signal, where the frequency spectrum of a signal is the frequency range that is contained by the signal. Mel spectrogram is basically a spectrogram that is represented in Mel scale.
We used the library Librosa in order to create Mel Spectrograms. In order to save them according to their class ID, we used metadata to match the filename with the class ID. The relevant Jupyter notebook is available here.
This steps consists of grayscaling, resizing and normalization of the spectrograms using OpenCV. Further, the data is splitted to train, test, and validation sets. The relevant Jupyter notebook is available here.
After preprocessing, the spectrogram looks like
Finally, CNN model is used for training. The model summary is given below:
| Layer (type) | Output Shape | Param # |
|---|---|---|
| conv2d (Conv2D) | (None, 64, 64, 32) | 320 |
| max_pooling2d (MaxPooling2D) | (None, 32, 32, 32) | 0 |
| conv2d_1 (Conv2D) | (None, 32, 32, 64) | 18496 |
| max_pooling2d_1 (MaxPooling2D) | (None, 16, 16, 64) | 0 |
| conv2d_2 (Conv2D) | (None, 16, 16, 64) | 36928 |
| flatten (Flatten) | (None, 16384) | 0 |
| dense (Dense) | (None, 64) | 1048640 |
| dropout (Dropout) | (None, 64) | 0 |
| dense_1 (Dense) | (None, 64) | 4160 |
| dropout_1 (Dropout) | (None, 64) | 0 |
| dense_2 (Dense) | (None, 10) | 650 |
Relevant Jupyter notebook is available here.