Home

Welcome to the classify-spectrogram wiki!

This project is all about extracting musical data from audio by using various algorithms to process spectrogram data.

The current and simplest algorithm is K-nearest neighbors, which requires computing the Euclidian distance between the sample and each of the training data. Thus, the performance will not scale very well for larger amounts of training data.

Another aspect of this project is to attempt to accurately classify multiple musical notes within a single time slice, i.e. identify chords in the spectrogram. The initial implementation of this is to look only at the fundamental frequency and the overtones of a specific musical note to determine the relative probability that note is being played. Initial results are promising, although a better thresholding system will be needed to tell when a note is actually being played, vs when no notes are being played but one happens to have more noise in the note's frequency bands and is showing the highest probability.