UrbanSound8K Sound Classification

This project was written to build a sound classification model on the UrbanSound8K dataset. The dataset contains a total of 8732 audio samples in 10 different audio classes recorded from different environments.

Requirements

To run this project you need to have the following requirements:

• Python 3.8+
• TensorFlow 2.7+
• NumPy
• Pandas
• Matplotlib

Dataset

The UrbanSound8K dataset can be downloaded from the UrbanSound8K website. This dataset has the following characteristics:

• Total Number of Sound Samples: 8732
• Sound Classes: 10 (for example, air_conditioner, car_horn, children_playing, dog_bark, drilling, engine_idling, gun_shot, jackhammer, siren, street_music)
• Duration of each Sound Sample:** Variable (average duration approximately 4-5 seconds)
• Sampling Frequency:** 44.1 kHz

Code

The script consists of the following sections:

• Data loading and preprocessing
• Model creation
• Model training
• Model rating

Data Loading and Preprocessing

The dataset was organized into folders representing each audio class. Each audio recording was converted into a grayscale image with a size of 128x128 pixels.

Model Building

The model consists of the following layers:

• 4 2D convolution layers
• 2 2D MaxPooling layers
• 2 Dropout layers
• 1 Flattening layer
• 1 full link layer with 128 units
• 1 full connection layer with 10 units

The model is trained using the Adam optimizer.

Model Training

The model is trained for 50 epochs. In each epoch, the dataset is sampled in a random order.

Model Evaluation

The model is evaluated on training and test datasets. The accuracy rate obtained on the training dataset is 99.5%. The accuracy rate obtained on the test dataset is 97%.

Result

The results show that the model is an effective sound classification model for the UrbanSound8K dataset.

Future Work

To improve the performance of the model, the following can be done in future studies:

• **The model can be trained further by using more epochs.
• A higher accuracy can be achieved by using a more complex model.
• The performance of the model can be evaluated on other datasets.