PCA Analysis on Voice Gender Dataset

Perform PCA (Principal Component Analysis) on the Voice Gender dataset to classify voices as male or female based on acoustic properties.

Dataset

The dataset consists of 3,168 recorded voice samples collected from male and female speakers. Preprocessing techniques using R packages seewave and tuneR were applied, analyzing the frequency range of 0Hz-280Hz (human vocal range).

PCA Analysis

Apply PCA to extract essential information from the dataset. Focus on the top 5 variables that capture approximately 80% of the information required for classification.

Model Comparison

Compare model performance on the entire dataset versus using only the 5 selected variables. Evaluate accuracy and F1-scores using the XGBoost classifier.

Results

The model trained on the entire dataset slightly outperforms the model using only 5 variables. However, the difference in performance is minimal, indicating that the selected variables contain significant classification information.

The entire dataset

x_train = train.iloc[:, :-1]
y_train = train["label"]
x_test = test.iloc[:, :-1]
y_test = test["label"]

model = xgboost.XGBClassifier()
classify(model,x_train,y_train,x_test,y_test)

              precision    recall  f1-score   support

      female     0.9730    0.9832    0.9781       476
        male     0.9830    0.9726    0.9778       475

    accuracy                         0.9779       951
   macro avg     0.9780    0.9779    0.9779       951
weighted avg     0.9780    0.9779    0.9779       951

The 5 selected variables

x_train2 = train[["meanfun","sd","median","Q25","Q75"]]
y_train2 = train["label"]
x_test2 = test[["meanfun","sd","median","Q25","Q75"]]
y_test2 = test["label"]

model = xgboost.XGBClassifier()
classify(model,x_train2,y_train2,x_test2,y_test2)

              precision    recall  f1-score   support

      female     0.9746    0.9664    0.9705       476
        male     0.9666    0.9747    0.9706       475

    accuracy                         0.9706       951
   macro avg     0.9706    0.9706    0.9706       951
weighted avg     0.9706    0.9706    0.9706       951

Conclusion

Using just the 5 variables can simplify the recognition model without sacrificing much in terms of performance.

Dependencies

• Pandas
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn
• XGBoost

Instructions

1. Clone the repository.
2. Install the required dependencies: pip install -r requirements.txt
3. Run the Jupyter Notebook to execute the PCA analysis and model comparison.