MOTIVATION
Websites like pandora.com and last.fm provide music recommendations based on the high level prosodic features as seen in the current song that you are listening to or by going through the music in your computer to make sense of your song preferences. You might notice that the Music Information Retrieval (MIR) algorithms used in these websites define and search for factors like major key tonality, acoustic instrumentation, simple harmonic progressions and so on if you ventured through these sites and clicked on the option "Why was this song chosen". In this project, I try to define, characterize one such parameter that is specifically and perceptually defined for Indian classical music but has no quantitative basis. Listeners of Indian classical music are very particular that they listen to a raaga like Darbari Kanada only in the night and a raga like Todi only in the morning. So, this could be a potential add-on to an application like pandora.
QUESTION
Indian classical music is very uninque in that it is spiritual in nature. It is believed that certain ragas are most beneficial when sung at an appropriate time during the day. This is referred to as the Time theory of Raagas. Can the reason for this compartmentalization be attributed to the rhythm strucure in the song ? This project tries to address this question by studying the rhythmic structure of the tabla performances in each of the songs. A word of caution: I am not trying to correlate the tala ( Rhythm cycle ) and the Raaga. But the idea is to analyze and check if the percussionist emotes in a way that has a relationship to the time theory of Raagas. Previous research in this topic has dealt with only melody of the song and was successfully able to quantify and demonstrate the time theory using a Harmonic Network.
DATA
The database I used for this analysis was borrowed from the ITC- Sangeet Research Academy. Ideally, I would have liked to create or use a database that gave emphasis only to the tabla performance of the song. I chose 4 songs from each category : Morning, Afternoon , Evening and Night. All the songs are set to teentala ( 4+4+4+4 = 16 beats/cycle ) except the song in Todi. This track is set to a teentala variant. We will discuss this song in detail later on. Analysis was performed on 60 second samples from each of these 16 tracks. Analysis data comprised of manually annotated onset instances of the tabla performances using Sonic Visualizer. Given a dataset where the tabla information can be viewed independantly, an automatic onset detection algorithm could be used. By Onset, I mean both the from the left and the right part of the tabla.These are referred to as the Bhayan and Dayan. The onsets would amount to between 200 and 400 strikes in any given minute of the song. Data also conisted of manually annotated modulation factor onsets. Some people refer to this as the "Gumiki". It is played in the left drum of the tabla and is a frequently used ornamentation technique. This data forms the second annotation variable. I will try to avoid using indian percussion terminology as much as possible. This is the link to the database.
METHOD
The parameters that are calculated from the annotation information are Instantaneous tempo, Histogram of Onsets, Histogram of Modulation factor and Windowed Normalized Pairwise Variability Index (nPVI). The Instantaneous Tempo is calculated as a function of Onsets per minute (opm) instead of the conventional beats per minute(bpm). The histogram of onsets represents the density of strikes in 30 equally spaced bins for the 60 second samples. The histogram of Gumiki is the density of modulation factor onsets in the bayan. The nPVI is a quantitative measure of the durational contrast. It is calculated for each 10 sample window in the tabla onset data of each piece. These parameters are computed using a program written in MATLAB. So do we see trends from plotting graphs for the these parameters ? In other words, do these graphs indicate a pattern that can be linked to the time theory of raagas. It is important to note that rhythm alone is definitely not the only factor that needs to be accounted for in this analysis. A rigorous model quantifying the time theory of raagas will involve the melody,rhythm and an interplay between the two. Having said that, we do notice some trends when rhythm alone is considered.
RESULTS - INSTANTANEOUS TEMPO
Onsets per minute Vs Instants plots
Does instantaneous tempo show any particular pattern for songs in Morning, Afternoon, Evening & Night ? The answer is No. Tabla performance speed depends on the song. You can have 2 songs of the same raga in very contrasting speeds. But what is very interesting to note is the very high range in the onsets per minute values. It is not common to note a four-fold increase in speed on a regular basis for the accompanying tempo. The graphs indicate that each track has a opm range between 200 and 800 ! This indicates that the activity factor in Indian percussion instruments is an important factor to analyse. This forms the motivation for the density plots in this project. The following figure shows the instantaneous tempo for the 16 tracks chosen.
RESULTS - HISTOGRAMS - ONSETS
Density Vs Bin number
Morning & Afternoon- There is relatively Flat graphs.Indicative of less activity variation; Density average around 10 onsets in every bin ( 30 bins ) ; Todi is an exception because of the Tala variation.Evening & Night- The Graphs are corrugated. Indicative of greater activity variation; Density averages slightly above 12 onsets in every bin .
RESULTS - HISTOGRAMS - GUMIKI
Density Vs Bin number
Notice the large peak in Todi due to the variation in teental which emphasizes on ornamentation. The two histograms establishes the fact that although there is no variation in the rhythm structure of the talas (4+4+4+4) we notice emphasis on gumiki and on activity variation. Gumiki density fluctuation seems to be low for Morning and afternoon songs. The fluctuation seems to increase for songs in the evening and night.
RESULTS - nPVI - 10 sample window
nPVI Vs window number
Overall decrease in nPVI noticed in all pieces. Morning and Afternoon have nPVI where in there is a tendancy to show sharper fall at some point in the song ; but we notice relatively smooth curves . In Evening and Night songs the decrease is not very sharp like above ; curves seem to be more corrugated
CONCLUSION
Categorization into 2 parts ie. Purva and Uttara rags can be established from the plots. Further categorization of each of these into 2 parts would probably require the combination of rhythm and melody. Regularity in the Histogram for onsets,Histogram for gumiki's,nPVI values and depiction of variation in teental using instantaneous tempo and the other plots are the key results of this project.