Deep Audio Classifier: An Artificial Neural Network Approach

network, Mel-frequency cepstrum coefficients, Librosa.

1. Introduction

2. Literature Review

Audio classification has been a subject of extensive research, given its critical applications in environmental monitoring, multimedia systems, and urban planning. Early approaches relied on handcrafted features and traditional machine learning models, such as SVMs and k-Nearest Neighbors (k-NNs) [8], for audio recognition tasks. These methods primarily focused on extracting features like Mel-Frequency Cepstral Coefficients (MFCCs), zero-crossing rates, and spectral roll-off to capture audio signal characteristics. While adequate for basic classification tasks, they struggled with more complex urban soundscapes where overlapping frequencies and noise posed significant challenges.

In recent years, deep learning has revolutionized the field of audio classification [9], [10]. CNNs have been extensively explored due to their ability to learn spatial hierarchies and localized patterns from spectrograms [11]. Studies have demonstrated that CNNs outperform traditional machine learning models in tasks such as music genre classification and environmental sound recognition, achieving higher accuracy by capturing temporal and spectral information [12], [13]. Similarly, ANNs, although less specialized than CNNs for spatial data, have shown promise in scenarios where computational simplicity and interpretability are prioritized. The UrbanSound8K dataset has become a benchmark in audio classification research, offering a diverse collection of labeled urban sound samples [14]. Researchers have used it to evaluate various architectures, such as Long Short-Term Memory (LSTM) [15] networks for sequential audio data and ensemble models combining CNNs with traditional classifiers. Data preprocessing techniques, including noise reduction, normalization, and MFCC extraction, have been consistently emphasized to enhance model performance by mitigating the effects of background noise and signal variability.

Despite these advancements, there remain challenges in accurately classifying urban sounds, mainly ambient and overlapping noise classes. Recent works highlight the potential of hybrid architectures and transfer learning to address these limitations. Integrating data augmentation and advanced feature extraction methods has also been proposed to improve the generalizability of models.

This study builds upon prior work by comparing the performance of multiple models—ANN, SVM, CNN— on the UrbanSound8K dataset. By leveraging MFCC-based features and rigorous preprocessing, this research aims to contribute to developing more robust and accurate audio classifiers, addressing current gaps in classifying complex urban soundscapes.

3. Methods

3.1 Preprocessing Steps

Librosa

Librosa is a Python music and audio analysis package. It provides the building blocks to create music information retrieval systems [16].

In the context of urban sound classification, it can extract features from audio recordings of city sounds, such as traffic noise, and then use those features to train a machine learning model to classify new audio recordings.

Mel-frequency cepstral coefficient

An MFCC comprises a number of coefficients known as MFCCs. They were created using an audio clip’s cepstral representation (A nonlinear “Spectrum-of-a-Spectrum”). The Mel-Frequency Cepstrum (MFC) differs from the cepstrum in that the frequency bands are evenly spaced on the Mel scale, which more closely resembles the human auditory system's response than the linearly-spaced frequency bands used in the conventional spectrum. When used in audio compression, this frequency warping can improve the representation of sound and potentially lower the transmission bandwidth and storage needs of audio signals. Feature extraction is a special form of dataset reduction. Using feature extraction techniques for extracting specific features from the speech, these features carry the characteristics of the particular speech, which help differentiate the different speech so that these features will play a significant role in speech recognition. Compressing a voice signal into streams of acoustic feature vectors, also known as speech feature vectors, is the first step in speech recognition. The idea of feature extraction is divided into two steps: 1) the speech signal is transformed into feature vectors, and 2) the useful characteristics impervious to changes in the surroundings and speech variation are selected. In speech recognition systems, however, where accuracy has drastically declined in the case of their existence, changes in ambient variables and variances in speech are significant. The MFCC features, the most popular and reliable due to their precise estimation of the speech parameters and effective computational model of speech, are unquestionably the most often utilized speech features [17], [18]. Fig. 1 shows the MFCC vs. time.

3.3 Model Architecture

3.3.1 Support vector machine

I. Type: Linear classifier II. Input: MFCC features extracted from audio samples III. Kernel: Linear (For simplicity) or Radial Basis Function (RBF) kernel to handle non-linear patterns IV. Output: Multi-class classification using the "one-vs-rest" strategy to differentiate the 10 urban sound classes

3.3.2 Random forest architecture

I. Type: Ensemble classifier II. Input: MFCC features extracted from audio samples III. Number of trees: 100 (Default setting) IV. Splitting criterion: Gini impurity or Entropy for node splitting V. Output: Majority voting among trees for class prediction

3.3.3 Artificial neural network architecture

I. Input layer: 40 MFCC features as input II. First dense layer: 100 neurons, ReLU activation, Dropout (0.5) III. Second dense layer: 200 neurons, ReLU activation, Dropout (0.5) IV. Third dense layer: 100 neurons, ReLU activation, Dropout (0.5) V. Output layer: Softmax activation with 10 neurons for multi-class classification.

VI. Optimizer: Adam VII. Loss function: Categorical crossentropy

3.3.4 Convolutional neural network architecture

I. Input layer: Spectrogram images of audio data II. Convolutional layers: Two 2D convolutional layers with ReLU activation (e.g., 32 filters of size 3x3) III. Pooling layers: Max-pooling layers after each convolutional layer for down-sampling IV. Fully connected layer: Dense layer with 128 neurons, ReLU activation V. Output layer: Softmax activation with 10 neurons for multi-class classification VI. Optimizer: Adam VII. Loss function: Categorical crossentropy

4. Experimental Setup

5. Experimental Results

6. Conclusion

This study demonstrates the effectiveness of different machine learning models in classifying urban sound data, with a particular focus on the UrbanSound8K dataset. We evaluated four models: 1) SVM, 2) RF, 3) ANN, and 4) CNN, and compared their performance in terms of accuracy, precision, recall, and F1-score. The results indicate that the CNN outperformed the other models with an accuracy of 87.50%, showcasing its potential for more complex, noisy data such as urban sounds. The ANN model followed with an accuracy of 80.20%, indicating that deep learning architectures are suitable for audio classification tasks. The RF and SVM models also performed well but with lower accuracies, emphasizing that feature learning in deep models may yield better results when dealing with the challenges of audio data classification.

This study highlights the value of deep learning approaches in audio classification, with CNNs standing out as the most effective model for this task. Future work can further explore data augmentation techniques, advanced hyperparameter tuning, and different feature extraction methods to continue improving classification performance.

Acknowledgments

We want to express our sincere gratitude to the contributors of the UrbanSound8K dataset, which has been instrumental in the success of this research. We also thank the developers of the machine learning libraries, such as TensorFlow, Keras, and Scikit-learn, for their extensive support in implementing and testing the models—special thanks to our faculty members and peers for their valuable feedback and assistance throughout this project.

Author Contributaion

Abhishek Raj: Led the project, designed experiments, and wrote the manuscript.

Abhishek Yadav: Handled data preprocessing, feature extraction, and model implementation.

Vineet Kumar: Implemented CNN and ANN models and analyzed results.

Abhay Kumar: Implemented SVM and RF models and visualized results.

Sankalp Anand: Assisted with data collection, literature review, and manuscript feedback.

Data Availability

The dataset used in this research, UrbanSound8K, is publicly available and can be accessed via the official website “https://urbansounddataset.weebly.com/urbansound8k.html”. The code and models developed during this research are available upon request.

Conflicts of Interest

The authors declare that they have no conflicts of interest related to the content of this research paper.