The code in this repository deals with the problem of predicting a speaker's gender, based only on the characteristics of their voice. We formulate the problem as a classification task between two classes (Male and Female), and we use Logistic Regression to perform the classification.
This task is part of an assignment in a Machine Learning (ML) course, and focuses on implementing a classification algorithm for a particular problem, evaluating its performance, and comparing it with any existing implementations.
The dataset we used in this implementation is freely and openly available in Kaggle. It consists of 3,168 recorded voice samples, collected from male and female speakers, and pre-processed to extract a number of acoustic features.
The dataset consists of 20 independent variables:
- meanfreq: mean frequency (in kHz); float
- sd: standard deviation of frequency; float
- median: median frequency (in kHz); float
- Q25: first quantile (in kHz); float
- Q75: third quantile (in kHz); float
- IQR: interquantile range (in kHz); float
- skew: skewness; float
- kurt: kurtosis; float
- sp.ent: spectral entropy; float
- sfm: spectral flatness; float
- mode: mode frequency; float
- centroid: frequency centroid; float
- meanfun: mean fundamental frequency measured across acoustic signal; float
- minfun: minimum fundamental frequency measured across acoustic signal; float
- maxfun: maximum fundamental frequency measured across acoustic signal; float
- meandom: mean of dominant frequency measured across acoustic signal; float
- mindom: minimum of dominant frequency measured across acoustic signal; float
- maxdom: maximum of dominant frequency measured across acoustic signal; float
- dfrange: range of dominant frequency measured across acoustic signal; float
- modindx: modulation index; float
1 target variable:
- labels: male/female (string)
To run the code first install the required dependencies:
pip install -r requirements.txt
Then you can run the classification task by running:
python main.py
The output consists of a set of calculated evaluation metrics printed in the console, and a number of plots, all illustrating the classification's performance.
In our code, we execute the following workflow:
- First we read the data from a csv file and split them into features and labels.
- Then we plot a data correlation matrix, in the form of a heatmap, that shows the pairwise correlation between the different columns of the data.
- Then, we encode the string labels (male and female) into numbers (0 for female, 1 for male).
- Then, we normalize the feature values to the range [0,1] using a MinMax scaler. This is necessary as the different features have different ranges that may negatively affect the classifier.
- Then, we split the data into train and test sets, with the train set consisting of 80% of the data and the test one of 20%. This splitting is random in every run.
- Then, we initialize a Logistic Regression model.
- Then, we define the scoring metrics that we want to evaluate the model with. These are accuracy, precision, recall, F1, and confusion matrix.
- Then we run and evaluate the classification model by performing a 10-fold cross-validation. We use cross-validation to ensure our evaluation is not biased from the random splitting of the data.
- Finally, we print the average values of the scoring metrics and plot the confusion matrix for each fold.
Our implementation can be compared with the implementation reported here. The main differences between them are the following:
- That implementation uses Support Vector Machines (SVM) as the classification algorithm and achieves an accuracy of 98.26%, while our implementation uses Logistic Regression and achieves an accuracy of 96.92%.
- That implementation reports only the accuracy score and a single confusion matrix, while our implementation reports also precision, recall and F1, and plots the confusion matrix per fold.
- That implementation standardizes the features by removing the mean and scaling to unit variance, while our implementation scales the features to the range [0,1].
- That implementation uses GridSearch to find the best parameters for the SVM algorithm, while our implementation does not do any parameter tuning.