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.
To set this project up, you need to proceed with the following steps:
- Install Conda for Python 3.8
- Clone this repository
- Run these commands to setup MFA
conda create -n aligner -c conda-forge openblas python=3.8 openfst pynini ngram baumwelchconda activate alignerpip install montreal-forced-alignermfa thirdparty downloadmfa download acoustic english
- Install the packages
flask,textgrid, andstringto setup the flask app - Run these commands to run
app.pycd flask_appexport FLASK_APP=appflask 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_appdirectory - Not running Conda for python 3.8
There are 2 pages total:
- Uploading page (http://127.0.0.1:5000/) - Upload your audio files and transcripts on this page.
- Play page (http://127.0.0.1:5000/play/) - Play your most recently uploaded audio file synced up with its transcript on this page.
A video demonstration is available here: https://youtu.be/dirdbHHzgwA
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