I have been working on a mentorship with Dr. Webber. I am trying to use RQA to segment music into meaningful sections.
I took the wave file of the music, cut its sampling frequency down to 5,000 Hz from 41,000Hz. (Wave files are numerical samples of the amplitude of the sound wave).
I put the data through RQA epochs (1500 points non-overlapped at 7% of max value radius and an embedding demension of 10) and found interesting results. Determinism was was an excellent marker for clear shifts in song sections and was almost always equal to the shannon entropy times a constant (parts where determinism/ shannon entropy spiked did not seem to have any clear meanng in the song though). I was also very surpised to find that a demension of 10 gave the best data.
it was also interesting to note that a statisical average smoothing of determinism gave a much better graph than using a larger epoch which would have covered the same area.
looking at a recuurence plot of Determinism showed interesting results. Lines that effectivly delineated the areas of singing showed up when compared to a short section of the song featuring the singer using heavy vibrato which happened a couple times throught the song.
I was wondering if anyone could explain some of these mysteries, offer constructive critisicism, or had encountered thign type of thing before. I am aware of previous research that used hidden markov models and was very effective.
I am interested in seeing if I could use RQA for beat detection by playing with delay, which I have sadly not really touched. I was also interested in doing cross recuurence plots of the determinism of two related songs to see if I coudl achieve a recognition of songs that were "similar" or sections that were similar. I have been using live and recorded versions of many songs but have had poor luck with this.
music segmentation
- Norbert
- Expert
- Posts: 194
- Joined: Wed Jan 4, 2006 11:03
- Affiliation (Univ., Inst., Dept.): Potsdam Institute for Climate Impact Research, Germany
- Location: Potsdam, Germany
- Location: Potsdam Institute for Climate Impact Research, Germany
As far as I know, Michael Casey has used a "similarity plot" (ixegram) in order to classify and compare songs. However, instead of using neighbourhoods of states (as in recurrence plots), he used other features of the data like locally estimated histograms (see
http://www.recurrence-plot.tk/related_methods.php and http://musicstructure.com)
A variation could be to build a classifier using the RQA measures, as you mentioned. A recurrence plot of RQA measures would be something like a meta-recurrence plot.
I would like to see such a recurrence plot computed from determinism. I'm afraid, that different playings of the same song may yield very different values of the RQA measures, even if their recurrence plots look similar.
Looking for the beat it could be helpful to compute the RQA measures diagonalwise as functions of the distance from the main diagonal (LOI), e.g. for the recurrence rate:
![Image](http://www.recurrence-plot.tk/1.pictures.recurrence/eq_RR_k2.png)
Maybe this would give you some hint about the beat intervals.
Norbert
http://www.recurrence-plot.tk/related_methods.php and http://musicstructure.com)
A variation could be to build a classifier using the RQA measures, as you mentioned. A recurrence plot of RQA measures would be something like a meta-recurrence plot.
I would like to see such a recurrence plot computed from determinism. I'm afraid, that different playings of the same song may yield very different values of the RQA measures, even if their recurrence plots look similar.
Looking for the beat it could be helpful to compute the RQA measures diagonalwise as functions of the distance from the main diagonal (LOI), e.g. for the recurrence rate:
![Image](http://www.recurrence-plot.tk/1.pictures.recurrence/eq_RR_k2.png)
Maybe this would give you some hint about the beat intervals.
Norbert