raw-audio-gender-classification

This project contains the code to train a gender classification model that takes raw audio as inputs.

The weights of the model from the article can be found in the models/ directory.

See my Medium article for more discussion.

Instructions

Requirements

Make a new virtualenv and install requirements from requirements.txt with

pip install -r requirements.txt

This project was written in Python 2.7.12 so I cannot guarantee it works on any other version.

Run tests

python -m unittest tests

Data

Get training data here: http://www.openslr.org/12

• train-clean-100.tar.gz
• train-clean-360.tar.gz
• dev-clean.tar.gz

Place the unzipped training data into the data/ folder so the file structure is as follows:

data/
    LibriSpeech/
        dev-clean/
        train-clean-100/
        train-clean-360/
        SPEAKERS.TXT

Please use the SPEAKERS.TXT supplied in the repo as I've made a few corrections to the one found at openslr.org.

Training

Run run_experiment.py with the default parameters to train the model with the performance discussed in the article.

Processing audio

Run process_audio.py, specifying the model and audio file to use. The audio file must be a .flac file.

This script makes many predictions on different fragments of the target audio file and saves the results to data/results.csv.

I used this script to produce the data for the video embedded in the Medium article.

Notebooks

I have uploaded two notebooks with this project.

Model_Performance_Investigation gives a breakdown of the performance of the model over the different speakers in the LibriSpeech dataset.

Interview_Segmentation is where I analysed the results of the process_audio.py script on an interview between Elton John and Kirsty Wark.