High-Dimensional Audio Genre Classification Using MFCC-Driven Multi- Layer Perceptrons

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025 www.irjet.net p-ISSN: 2395-0072

High-Dimensional Audio Genre Classification Using MFCC-Driven MultiLayer Perceptrons

Chiranjeevi Vantaku

Department of Computer Science & Electronics and Communication Engineering Maharaj Vijayaram Gajapathi Raj College of Engineering (MVGRCE), Vizianagaram, Andhra Pradesh, India

Abstract - Audio data, as a form of unstructured information, presents both opportunities and challenges in the field of machine learning. Classifying such data effectively requires thorough preprocessing to transform rawsignalsintomeaningfulfeatures.Thisprojectfocuseson building a music genre classifier using artificial neural networks, specifically a multi-class model capable of predicting one of ten genres: blues, classical, country, disco, hip-hop, jazz, metal, pop, reggae, and rock. The goal is to develop a probabilistic model that not only identifies the most likely genre but also provides insight into the confidenceofitspredictions.

Key Words: Audio Classification, MFCC, Multilayer Perceptron, Deep Learning, Music Genre Recognition, Neural Networks, Signal Processing, Feature Extraction, PCA, Overfitting Mitigation

1.INTRODUCTION

Audioclassificationhasbecomeanincreasinglyimportant area within machine learning, with applications spanning diverse industries. It sits at the intersection of signal processing, machine learning, and acoustics, enabling systems to interpret and categorize sound data. From detecting speech and analyzing environmental noise to recognizing emotional tone in human voices, audio classification supports a wide range of real-world use cases and continues to grow in relevance as audio data becomesmoreprevalent.

1.1 The Dataset

A critical foundation for effective audio classification lies in selecting a dataset that accurately represents the problem space. For music genre classification, an ideal dataset should maintain consistent audio quality, standardizedfileproperties,andwell-definedgenrelabels. The GTZAN Genre Collection, curated by George Tzanetakis, meets these requirements and serves as the primary dataset for this project.It consists of 1,000 audio tracks, evenly distributed across 10 distinct genres, with 100 tracks per genre. Each track is a 30-second .WAV file in lossless format, sampled uniformly at 22,050 Hz. This standardized sampling rate ensures reliable feature extraction across all tracks, minimizing discrepancies due toover-orunder-sampling.

1.2

Preprocessing Methodology

Machine learning algorithms cannot directly interpret raw audio data due to its unstructured nature. To bridge this gap, the audio signals must first be transformed into structured numerical representations. In this project, the raw audio files were processed using signal transformation techniques to extract Mel Frequency CepstralCoefficients(MFCCs) awidelyusedfeatureset for audio classification tasks. MFCCs are particularly suited for music genre recognition because they reflect how humans perceive sound frequencies, emphasizing lowerfrequencies wherethehuman ear ismost sensitive. This perceptual alignment makes MFCCs highly effective for applications where subjective interpretation of sound playsarole.

2. Methodology

2.1

Technologies Used

ArangeofPython-basedlibrarieswereemployedtobuild andevaluatethemusicgenreclassificationmodel.NumPy andPandasservedasthecorenumericalengines,enabling efficient handling of linear algebra operations and highdimensional data structures. For audio preprocessing, the Librosa library was utilized to extract relevant audio features, including MFCCs. Visualization and performance tracking were conducted using Matplotlib, providing clarity on model behavior and data trends. Scikit-learn

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025 www.irjet.net p-ISSN: 2395-0072

facilitated the splitting of the dataset into training and testing sets, ensuring a structured pipeline. Finally, the classification model was constructed using Keras with a TensorFlowbackend,offeringahigh-levelAPIforbuilding andtrainingdeeplearningmodels.

2.2 Preprocessing Implementation

The initial step in building an effective music genre classifier involves standardizing the dataset to ensure consistent input for the machine learning model. This requires systematically extracting meaningful features from each audio file across all genre categories. For each segmented window of an audio track, Mel Frequency Cepstral Coefficients (MFCCs) were derived using a sequence of signal processing techniques, including Fast FourierTransforms(FFT),Short-TimeFourierTransforms (STFT), and log-scaled spectrogram analysis. These methods collectively transform raw audio signals into structured representations that are suitable for training themodel.

MFCCs capture key characteristics of an audio signal across three dimensions: frequency, magnitude, and time. Inthisproject,eachMFCCwascalculatedusing13distinct coefficients, a standard choice in audio analysis tasks due to its balance between complexity and representational power [9]. These coefficients are stored in a 13dimensional array, where the array’s length corresponds tothenumberofMFCCwindowsderivedfromeachaudio track.Thetotalnumberofsamplespertrackisdetermined by multiplying the audio sample rate by the track’s duration,ensuringconsistentinputdimensionsacross the dataset.

(22, 050Hz × 30sec) = Samples / track (1)

An additional benefit of using MFCCs over log spectrograms lies in their reduced dimensionality. While log spectrograms can produce thousands of features per audio sample, MFCCs are confined to just 13 predefined coefficients. This significant reduction in feature space makes MFCCs far more efficient for computationally demanding tasks like model training. Lower dimensionalitynotonlydecreasesmemoryusagebutalso helps in reducing training time and mitigating the risk of overfitting,therebyimprovingoverall model performance andscalability.

2.3 Exploratory Data Analysis

After converting the raw audio signals into structured numerical features, the dataset becomes suitable for modeling and pattern recognition. Prior to training, the datawasdividedintotrainingandtestingsetsusingScikitlearn to ensure unbiased evaluation. Given the highdimensional natureofthe extractedMFCCs – eachsample forming a matrix of size 130 × 13 – dimensionality reduction was applied to facilitate exploratory data analysis.PrincipalComponentAnalysis(PCA)wasusedto project the data into a lower-dimensional space. The resultingscatterplotsrevealedthatthedataisnotlinearly separable, which influenced the decision to adopt more

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025 www.irjet.net p-ISSN: 2395-0072

complex, non-linear classification models such as neural networks.

Figure 5: PCA 2D Decomposition

2.4

Model Selection

Due to the high dimensionality of the input features and the need to generate a multiclass probability distribution, a neural network architecture was chosen for this classification task. Specifically, a Multilayer Perceptron (MLP) was implemented using Keras, taking advantage of the flexibility and learning capacity of artificial neural networks (ANNs). The model architecture included four hidden layers, trained using the backpropagation algorithm to optimize weight adjustments based on error feedback. The layer configuration consisted of an input layer with 1690 units, followed by hidden layers of 1024, 512, 256, and 64 neurons respectively, and a final output layer of 10 nodes representing the target music genres. Model performance was evaluated using sparse categorical cross-entropy as the loss function, along with accuracymetricsonbothtrainingandvalidationdatasets. These measures were instrumental in diagnosing overfitting and fine-tuning model parameters for improvedgeneralization.

3 Machine Learning Techniques

3.1

Computational Complexity

To minimize computational complexity and reduce the training time during parameter tuning, a simplified version of the dataset was employed. This reduced set retainedthestructure of theoriginal dataset but included onlyonerepresentativesongpergenreinsteadof100.By limitingthedatainthisway,theoveralltrainingtimewas decreased by a factor of 100, allowing for quicker experimentation and model iteration during the optimizationphase.

3.2 Parameter Configuration

To streamline parameter optimization and reduce training time, a scaled-down version of the dataset was utilized.Whilestructurallyidenticaltothefulldataset,this reduced set contained only one audio track per genre instead of 100. This strategic reduction significantly lowered computational demands, cutting training time by approximately 100-fold, and allowed for faster testing of modelconfigurations.

Following the initial evaluation of model performance on both the training and validation sets, several adjustments were made to enhance learning efficiency. One key modification involved tuning the learning rate, as the modelwasconvergingtoorapidly,potentiallyoverlooking optimal solutions. To address this, the default Keras learning rate of 0.01 was reduced to 0.0005 through empirical testing. This lower learning rate slowed down convergence, allowing for more gradual and stable training. Additionally, increasing the number of epochs provided more opportunities for weight updates, leading to smoother and more interpretable learning curves. This adjustment helped mitigate the issue of overly steep loss andaccuracytrends,makingthetrainingprocesseasierto monitorandrefine.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025 www.irjet.net p-ISSN: 2395-0072

3.3 MLP Optimization

3.3.1

Hyper Parameterization

Withthelearningrateappropriatelytuned,thenextfocus was on addressing the gap between training and validation performance, particularly to assess the model’s ability to generalize to unseen data. A noticeable discrepancy,wherevalidationlossexceedstrainingloss,is a clear indicator of overfitting. To mitigate this, regularization techniques were applied. Specifically, L2 regularization (also known as ridge regression) was introduced using a kernel regularizer, which penalizes large weight values to encourage simpler models. The regularization strength, denoted by λ (lambda), was initiallysetto0.01.However,afterevaluatingconvergence patternsusinglossandaccuracyplots,theoptimallambda value was found to be 0.001, effectively balancing model complexityandgeneralization

3.3.2 Dropout Layers

BeyondL2regularization,additionalmeasuresweretaken to further reduce overfitting. Dropout layers were introduced after each of the three dense layers to randomlydeactivatea portionofneuronsduringtraining, promoting robustness in the network. The default Keras dropout rate of 0.3 was reduced to 0.2, which yielded better performance when used alongside the L2 regularization term. To complement these efforts, early stopping was also implemented as a final safeguard. With themodelsettotrainforupto100epochs,earlystopping allowedtrainingtohaltautomaticallyifthevalidationloss ceasedimproving,therebypreventingunnecessaryweight updatesandreducingtheriskofoverfitting.

3.3.3 Activation Functions

Given the structure of artificial neural networks (ANNs), one critical challenge that arises during training is the vanishing gradient problem, particularly when using the backpropagation algorithm. This issue is common with traditional activation functions like sigmoid or hyperbolic tangent, whose gradients approach zero at extreme input values, leading to stalled learning in deeper layers. To address this, the Rectified Linear Unit (ReLU) activation functionwasappliedtoallhiddenlayers,asitmaintainsa stronger and more consistent gradient during training. The softmax function was reserved solely for the output layer,whereitwasusedtoproducea10-classprobability distributioncorrespondingtothetargetmusicgenres.

4 Results

4.1 Classifications

After applying various optimization techniques to minimizeerrorandoverfitting,themodelachieveda64% accuracy on the validation dataset. This result is considered promising, especially given the model’s probabilistic output design. Unlike single-label predictions, a multiclass probability distribution offers deeperinsights byindicatingthelikelihoodof eachgenre. Forinstance,ifasongismisclassified,itisstillvaluableto observe that it shows high probabilities across related genres like pop, country, and rock, rather than receiving only the top predicted label. This richer output aids in understanding the model’s behavior and the inherent overlapbetweenmusicgenres.

4.2 Cross Class Confusion Matrix

The confusion matrix shown in Figure 9 supports this observationbyvisuallyillustratingpatternsinthemodel’s predictions. A strong diagonal trend indicates high accuracy in correctly identifying individual genres,

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025 www.irjet.net p-ISSN: 2395-0072

suggesting solid single-class performance. In contrast, the lighter, off-diagonal regions reveal where misclassifications commonly occur highlighting genre pairs that the model tends to confuse. This visualization provides valuable insights into both the strengths and limitationsoftheclassifier.

Theconfusionmatrixalsoshedslightonthegenre-specific performance of the model. It clearly indicates that metal and classical music genres are classified with the highest accuracy, reflecting strong feature separation and consistent patterns within these classes. On the other hand, genres such as country, disco, hip hop, and rock exhibit comparatively weaker performance, likely due to overlapping audio characteristics and broader variability within those categories. This insight is valuable for identifying which genres may benefit from additional preprocessingorfeaturerefinement.

Conclusion

ThisstudydemonstratestheeffectivenessofusingMFCCs as structured features and MLPs as predictive models for audio genre classification. By focusing on dimensionality reduction, regularization techniques, and probabilistic output interpretation, we achieved a meaningful classificationaccuracyof64%onthevalidationset.While certain genre pairs posed classification challenges due to overlapping acoustic features, the model showed strong performanceonmoredistinctgenressuchasclassicaland metal. Future improvements involving deeper architectures, ensemble learning, and metadata integration could enhance performance further. The findings underscore the potential of deep learning in extracting and modeling patterns from high-dimensional unstructuredaudiodata.

REFERENCES

1. GeorgeTzanetakisandPerryCook.MusicalGenre Classification of Audio Signals. IEEE Transactions onSpeechandAudioProcessing,10(5),2002.

2. Logan, B. Mel Frequency Cepstral Coefficients for Music Modeling. Proceedings of the International Symposium on Music Information Retrieval (ISMIR),2000.

3. Deng, L., & Yu, D. Deep Learning: Methods and Applications. Foundations and Trends® in Signal Processing,7(3–4),197–387,2014.

4. McFee, B., Raffel, C., Liang, D., Ellis, D.P.W., et al. Librosa: Audio and Music Signal Analysis in Python.Proceedingsofthe14thPythoninScience Conference (SciPy), 2015, https://librosa.org/doc/latest/index.html

5. Chollet, F. Keras: Deep Learning for Humans, https://github.com/keras-team/keras