Forced Aligner

This project hooked up Flask with the Montreal Forced Aligner (MFA) to create a webpage that, when given a .wav audio file and a .txt transcript, plays the transcript synced up with the audio.

Setup

To set this project up, you need to proceed with the following steps:

1. Install Conda for Python 3.8
2. Clone this repository
3. Run these commands to setup MFA
   • conda create -n aligner -c conda-forge openblas python=3.8 openfst pynini ngram baumwelch
   • conda activate aligner
   • pip install montreal-forced-aligner
   • mfa thirdparty download
   • mfa download acoustic english
4. Install the packages flask, textgrid, and string to setup the flask app
5. Run these commands to run app.py
   • cd flask_app
   • export FLASK_APP=app
   • flask run

The webpage can now be found at the address http://127.0.0.1:5000/.

Some mistakes that I've made in the process (that you should check if it isn't working on your end):

• Not being in the alignerconda environment
• Not being in the flask_app directory
• Not running Conda for python 3.8

How to use the webpage

There are 2 pages total:

1. Uploading page (http://127.0.0.1:5000/) - Upload your audio files and transcripts on this page.
2. Play page (http://127.0.0.1:5000/play/) - Play your most recently uploaded audio file synced up with its transcript on this page.

Video demonstration

A video demonstration is available here: https://youtu.be/dirdbHHzgwA

Next Steps

There were a lot of areas that could be improved (this is my first time using MFA and Flask), here are what I believe to be the lowest-hanging fruit that would increase performance significantly:

• Training custom, more extensive models using MFA. In the 2nd and 3rd test, rare/unclear words were frequently missed and it would be nice to catch them instead of having to estimate it.
• Designing a better estimation algorithm for unknown words