In this folder you can find the three best models as reported in the original paper. Here it follows a brief description on their main characteristics and how to load them in your scripts/notebooks.
This a 1-dimensional Convolutional Neural Network trained for the age regression task in this notebook: as input features it uses MFCC compute as described in Section 02 of this README file.
This models has been traind with Keras, therefore you can load them with the following instructions:
from tensorflow import keras
model = keras.models.load_model('age_regression-mfcc/')MAE: 9.443
This model is a classical Linear Regression model that takes as input i-Vectors. This model has been trained with Scikit-learn and can be loaded as follows:
import pickle
with open('age_regression-ivec-lm_unbalanced.pkl', 'rb') as f:
model = pickle.load(f)MAE: 9.443
This model, instead, is the one that obtained the highest F1-Score in the gender recognition task. It is a logistic regression model implemented in PyTorch, therefore you can load it with the following instructions:
from src.gender_classifiers import LogisticRegression
import torch
model = LogisticRegression(512, 1)
model.load_state_dict(torch.load('ivec_log_reg_model.torch'))
IMPORTANT: You will need to import this custom file containing all the models used for predicting gender.
F1-Score: 0.9829