Automatic Age Labeler for Voice Datasets

When creating machine learning models that incorporate voice, the age of the speaker can be a very useful feature for downstream voice tasks. Categorizing the age for unlabeled datasets can be very useful for creating models.

This application provides users an easy way of automatically labeling their voice samples with the age category of the speaker. The user simply needs to attach mp3 files, and press "upload" to download a CSV file with two columns: the filename, and the age category.

This repository contains notebooks that detail the entire machine learning project, including data preprocessing, exploratory data analysis, and modeling. It also contains code for the frontend/API, which can be run using Docker.

Directory Structure

├── ML-modeling                      # Notebooks with Keras models/analysis
├── data-prep                        # Data Preprocessing
├── eda                              # Basic Exploratory Data Analysis
├── presentation-slides              # Presentation from NeuroLex demo day
├── result-reports                   # Result Reports from 2018
├── webapp                           # Code for Web Application
    ├── client                       # client-side code
    ├── nginx                        # Reverse proxy for routing
    ├── server                       # Backend code                
├── .gitignore                       # Only push necessary files                 
├── Dockerrun.aws.json               # JSON for multi-container AWS deployment, not currently deployed
├── README.md                        # Readme that details the repo
├── docker-compose.yml               # Necessary to run the app on user machine

Running the Application with Docker

docker-compose up --build
# now the app is running on localhost

Update:

PyPI no longer supports building the dependencies that are needed to run the backend, (numpy==1.19.5, scikit-learn==0.21.3, tensorflow==2.7.0). To run this container, you will need to install Tensorflow 2.7 from the source within your virtualenv or Docker container.

References

• Jim Schwoebel's Voicebook

Demo

Further directions

If you are interested in getting involved with machine learning for audio analysis, I recommend you check out NeuroLex Labs, which provides voice computing services to clients, including machine learning. A special thanks to Jim Schwoebel, the CEO of NeuroLex, for giving me direction and help with this project!

Behavioral Signals is another interesting company that detects emotion in voice with AI. The Director of ML at the company, (Theodoros Giannakopoulos) developed pyAudioAnalysis, which is the library I used to featurize the audio data.

Technologies Used