WIT: Gender & Age Recognition

Who Is Talking (WIT) is a project aims to classify audio input and predict the gender, and age group using machine learning techniques.

Project Overview

We built a pipeline that takes in voice data and classifies:

• Speaker Gender (Male / Female)
• Age Group (Twenties / Fifties)

Installation

git clone https://github.com/Mohamed-Ashraf273/Who-Is-Talking.git
cd Who-Is-Talking
mkdir data
python -m venv venv
source venv/bin/activate 
pip install -r requirements.txt

How to Prepare Your Data for Training

If you plan to train the model on your own dataset, follow these steps:

1. Create a data_train/ directory in the root of the repository.
2. Place your audio files inside one or more subdirectories under data_train/, and make sure each folder name starts with audio (e.g., audio_batch1, audio_train_set, etc.).
3. It must also contain a .csv file that holds the corresponding labels for the audio files.

Example Directory Structure

Who-Is-Talking/
├── data_train/
│   ├── audio_batch1/
│   │   ├── sample_001.wav
│   │   ├── sample_002.wav
│   │   ├── labels.csv
│   ├── filtered_data_labeled.tsv
├── model/
│   ├── model.pkl
│   │── scaler.pkl
├── src
│   infer.py
│   main.py
│   pipeline.py
├── README.md

To train your own data run:

./wit train --features './path_to_features.csv' --model model_selected

1. model by default is "stacking".
2. if you want to grid search over "xgboost" or "lgbm" add "--grid_search" when running the command, grid_search is supported for "xgboost" and "lgbm" only.
3. for saving files like "test_val.json" or "test_data.json" add "--save_val" for test_val and "--save_test" for test_data.

To Extract features run:

./wit features --datapath './data_train/audio_batches' 

1. your batches name should start with "audio".
2. in your audio_batches dir you should include "filtered_data_labeled.tsv" contains your data.

To get performance run:

first check that the models in the model path.

./wit validate --testfile './test_val.json' --model model_selected --val

1. your validation file should be a json file.
2. model by default "stacking".

To get predictions file run:

first check that the models in the model dir.

./wit predict --audiopath './path_to_the_dir_contains_audios'

For inference you could also build our docker image then run it:

before you build the image make sure that the test path dir in infer.py is setted correctly.

docker build -t wit .
docker run -v "C:\path_to_the_project:/app" wit

Note: make sure you installed docker

OR
run the infer.exe file, but make sure that the test audios are in ./data dir.

Note: for developers

we are on 'black' format.

pip install black
python black ./path_to the_file