Author: gyl2

This week has been continuing to fine tune the algorithm for slurred notes and rests. We made the switch last week to STE from RMS and we saw better results as it kept the near zero values closer to zero, however with slurred notes the values still do not get close enough.

I created an algorithm that would look through the segments from the original algorithm and if it has any times, then it would “flag” the segment as a possible slurred note. Then I would check for pitch changes. I first used spectrogram but found that iw would miss some notes or incorrectly identify where the notes actually change, so I switched to using STFT and having a ratio varying based on BPM to detect smaller note changes with faster tempos with better success. Here is a picture of using the spectrogram

and using the STFT

While there are still some inaccuracies, it is much better than before. Currently working on fine tuning the rest detections as right now it has a tendency to over detect rests. Also looking into using CNNs for classification of slurred notes

While implementing this project, I learned more about signal processing and how some things are much easier to identify manually/visually than coding it up. Additionally, reading up on how much people research into identifying segmentations in music and how different types of instrument can add to more slurs as they tend to be more legato. For this project, I needed new tools on identifying new notes like using STFT and STE. To learn more about these, I would read research papers from other projects and universities on how they approach it and tried to combine aspects of them to get a better working algorithm.

Worked on the issues mentioned during interim demo, which were not being able to accurately detect slurs and not picking up the rests in songs.

First, experimented with modifying the code to Short Time Energy (STE). This helped the code become more “clear” as there were less bumps and more clear near zero values, essentially eliminating some of the noise that stayed with RMS. Should make amplifying the signal a lot easier now. However, still having some difficulty seeing the differences in slurred notes, so doing some additional research in onset detection to detect slurred notes specifically rather than look for the separation of notes in segmentation.

 (forgot to change label for the line, but this is a graph for STE – this was taken using audio from a student in the school of music)

For rests, modified my rhythm detection algorithm to instead look for zero values after reaching the peak (means the note is done playing) and taking the additional length after the note to count as a rest. Sometimes takes slight moments of silence to distinguish notes as a rest though so need to do some experiments to make it less sensitive.

This week I worked on finishing the rhythm detection. I approached this by using the audio segmentation code that I created last week. I looped through the segments and then in each segment, we use the BPM (which will come from the web app once it is integrated) to calculate the length of the notes and then using an if statement to classify it as either a sixteenth, eighth, quarter, etc note. This seems to be working with the Twinkle Twinkle Little Star audio. May be a little buggy with how I am calculating rests as I am just calculating the remaining portion of the segment so will need to figure out a better algorithm for this and test it further. Will be looking into using Regions of interest/energy to do audio segmentation for audios with less steep increases in amplitude (slurred note) for more precise audio segmentation. Currently on schedule – working on interim demo presentation and integrating shivi, deeya, and i’s part.

This week was dedicated to testing the audio segmentation code that I created last week. From here, I realized that I would need to change the peak difference threshold, or how steep a change must be to be considered a new note, as well as the max time difference, how long it must take for a note to be considered a new note, change with the BPM and audio quality. After testing with a random audio sample from youtube of Ten Little Monkeys, we realized that this audio was probably not that useable, at least for early stage testing, as the eighth notes were note distinct enough for the code, and quite honestly myself manually, to identify.

In this image, using the audio recording of ten little monkeys, I have manually identified where the notes changes using the green line. You can see here that the code (in the red lines) are not correctly identifying these notes. However, the change is not that significant either (less than a 0.3 RMS difference) and doesn’t approach zero as much as other audios do, like twinkle twinkle little star. To try and fix this, I experimented with different audio amplifications – first just scaling the signal, but this resulted in the “near zero” value also getting scaled, meaning it still wouldn’t pick up the note, though it had a significant difference now. Then I tried to scale it exponentially, to keep the near zero values near zero and the other values increase, but since it is fairly quiet overall, this meant that everything was near zero and didn’t have a significant difference. This led us to first experiment with clearer cut audios before coming back to this one when the program is more advanced.

My next steps would be finding more distinct/staccato recordings of faster notes and seeing how the audio segmentation handles that and polish off my rhythm detection code.

This week, I got audio segmentation up and working. After our previous conversation with Professor Sullivan, I first converted the audio signal into RMS values. 

My first approach was to calculate if there was a sharp increase in the RMS. However, this caused me to incorrectly identify some spikes multiple times. Increasing the amount of time that has passed since the last identified point often caused me to miss some beginning of notes. 

(image where the dots would signify the code identifying the start of a note, as you can see, it was too much)

I then realized that before notes, oftentimes the RMS would get near zero. So my next approach was to convert my code to identify when the RMS is near zero, but then when I got in a moment of silence (like a rest) I would incorrectly want to segment the silence into many different segments, which would waste a lot of time. So I tried to do a combination of the two then where I would look for when the RMS was near zero and then look for the nearest peak. Then if this nearest peak RMS minus the starting RMS (near zero) difference was greater than a specific threshold, currently 0.08 but this is still getting experimented with, then I would identify it as a correct note. While this was the most accurate approach thus far, I still ran into a bug where even in the moments of silence it would find the nearest peak, a couple of seconds away, and identify the silence as multiple beginning of notes again. I fixed this by checking how far away the peak was and making a maximum threshold. 

(where the dotted blue line would be where the code identified the nearest peak and the red line is where the code is marking the near zero RMS values)

Currently this works for a sample of twinkle Twinkle Little Star. When testing this with a recording of ten little monkeys, it works if we lowered the RMS threshold, which signified that we would need to standardize our signal somehow in the future. We also noticed that with quicker notes, the RMS values don’t get as close to zero as quarter or half notes, so we might need to increase the threshold for what is considered near zero. 

(red line is where code has identified beginning of notes and blue dotted line is where i manually identified the beginning of notes)