Speech Accent Classifier

Play around with the model and make predictions using the web app I made using Flask.

Read the blog post.

Description

Classifying audio of human speech into various accents/countries of origin using MFCC coefficients extracted from audio .wav files.

The model in flask_app/static/sklearn_models/final_model.pkl is an ensemble of K-Nearest Neighbor and Logistic Regression models. The overall predictive accuracy of the model is 0.89 and it has an ROC AUC score of 0.95. The blog post about it is here.

This was developed over a 2-week span in August 2020 as a project for the Metis data science program.

Data Source

Fokoue, E. (2020). UCI Machine Learning Repository - Speaker Accent Recognition Data Set. Irvine, CA: University of California, School of Information and Computer Science.

File Contents

• Running main.py will launch the Flask app locally.
• utilities.py contains files needed for the Flask app to run.
• notebooks/ contains the Jupyter Notebook used to do all data analysis and modeling, as well as an accompanying file of Python functions.
• templates/ contains HTML files for the Flask application
• static/ contains static files for the Flask application as well as the pickled scikit-learn model.
• heroku.yml and Dockerfile are used for deployment of the Flask app to Heroku

Dependencies

The contents of /notebooks can't be fully run because the Jupyter Notebook connects to a remote SQL database. In order to install the dependencies for the flask app, run:

pip install -r requirements.txt