How to determine if the extracted features are correct

fraunhoferportugal / tsfel

An intuitive library to extract features from time series.

https://tsfel.readthedocs.io

BSD 3-Clause "New" or "Revised" License

945 stars 143 forks source link

How to determine if the extracted features are correct #122

Closed Akai-ai closed 1 year ago

Akai-ai commented 2 years ago

When I extract the MFCC and LPCC feature matrix, how can I tell that there is no problem with my extraction? Is it possible to reverse the reconstruction to calculate their relative errors?

dmfolgado commented 2 years ago

Hi! All the features we have implemented are unit tested. You can convert MFCCs back to the waveform audio signal using the librosa.feature.inverse.mfcc_to_audio. As a naive approach to measuring the reconstruction error, I guess you can calculate the RMSE and/or MAE between the original and reconstructed waveforms.