Accenter

Build and Run Specifications

Use the requirements.txt file to properly set up Python virtual environment with pip install -r requrements.txt.

Directory Setup

For functions in dataUtil to work as intended, the following directory setup is recommended.

.
├── ...
├── data
│   ├── processed
│   │   ├── american
│   │   │   └── clips
│   │   ├── chinese
│   │   └── ...
│   └── raw
│       ├── american
│       ├── chinese
│       └── ...
└── ...

where each of american, chinese, british, etc. are considered "labels" for the classification task and contain subdirectories clip within each of their processed folders.

Dependencies

pydub throws warnings without an installation of ffmpeg, tools for manipulating multimedia files. Visit the pydub repository to properly set up ffmpeg.

Use pipreqs . to generate requirements.txt based on all imports with versions used in import statements in the project.

Classification

We implemented three models for classifying our three accent classes. All three can be found in models/classification. We trained the models on 15 epochs of our data and received results that can be seen in our final report.

Preprocessing

We first split audio data into 1 second segments, cutting out the non-speaking time as well. To do this, run python main.py segment <raw_data_dir> <output_data_dir> where raw_data_dir is the directory your raw audio files are in, and output_data_dir is the directory the segmented pieces will be extracted to with a matching internal file structure.

For the final model, we run an MFCC preprocessing step to extract features instead of training on the raw audio wave files. To do this on a set of data setup as show above in directory setup, run python main.py fextr <processed_dir> where processed_dir is the directory with all of the segmented audio clips.

Web Application

A simple Flask web app was built to demo Accenter with custom audio recordings. To start the Flask app, run

export FLASK_APP=app.py
flask run

to start the application locally.