Final update/devlog !!!!

I finally finished integrating the AI pronunciation system into Anukrama.

This ended up being way more complicated than I originally thought. The model itself was trained in PyTorch using MFCCs, spectral features, and DTW comparisons against multiple good chanting recordings. The problem was that the .pt model couldn’t just be dropped into a website, so I had to export everything to ONNX and basically rebuild the preprocessing pipeline in JavaScript so the browser could generate the same kind of audio features from a recording.

I also changed how I stored the scaler data by putting the normalization values into the .npz files, then wrote the JavaScript to load the model, scaler, reference audio, process the recording, calculate the features, and pass everything through the ONNX model. There were a bunch of stupid issues along the way too, including the ONNX runtime conflicting with a variable in my JS and the scaler data not loading correctly lol.

After all of that I added the actual score display to the site, so the recording can go through the whole pipeline and come back as a pronunciation score out of 100.

This took way longer than I expected lol, but I finally got it working.

This is prob the end of this project!!!!