Transcription of mridangam strokes, motifs of ragas

[hema]

I am reposting it in this forum -- apologies. This is primarily owing to my desperation to find someone.
Dear All
i) We are building machine learning algorithms to automatically transcribe tanis. As ML algorithms require ground truth during training we need personnel (chennai based) to help us transcribe data. I would be especially happy if the persons that can transcribe have some background in engineering. It would be even better if we can get an aspiring musician with above 65% in his/her basic degree, a GATE score in ECE or CSE and is also interested in academics. We will train him/her in ML and s/he can in turn teach us how to appreciate tani. The project is on until 2016, the person can therefore even qualify himself with an MS.
If anybody in this forum knows somebody who can help us, I would be very happy to interact with the person.

ii) We are also looking for an engineer cum musician who can verify work done raga motivic analysis. We need somebody who can vet the work that we are doing. Again it would be nice to have somebody with a background in engineering and is also a performing musician (same qualifications as in i))

Rgds, Hema
hema

Posts: 47
Joined: 20 Dec 2010 23:28
Top

[Nick H]

It is a very difficult thing. I don't know if it is any easier for tabla, but I am given to understand that there is more of a regular relationship between the spoken and the struck tabla bols. With mridangam, it varies from school to school, and even from lesson/composition to lesson. There is also a considerable variety of sound produced by the instrument. For instance, many artists believe that the ringing chapu sound is the essential sound of the mridangam, and cultivate it, while others do not, and the stroke may not sound that different to dhi.

For some reason, many rasikas who take a great interest in the intricacies of raga and song take less in the mechanics of playing an instrument, and even less in the mridangam. Unless they play, or have been students themselves. Many here, although quite familiar with the calculations of kalpana swara, are somehow lost in the calculations of the tani.

My guess is that you need to work with an accomplished mridangist on this, and the first phase would be establishing if such a thing is even possible. We have two or three pro mridangists amongst our membership who visit occasionally, but you can't rely on this site as a method of contacting them, especially as you need is urgent.

It would be great if such a thing were possible. Maybe one day there might even be an app! How wonderful it would be for students and enthusiasts to be able to read off the classroom lesson or tani that they just heard. I feel that that is a vain dream --- but I do not know nearly enough about either mridangam or programming to be sure.

[hema]

Nick
There is some work that Prof. C V Raman did a while ago using sand figures and connecting them to the physics of sound production. We were able to repeat the same using signal processing. I have Akshay Anantapadmanabhan working with me.
As you rightly pointed out there are issues about the strokes itself. To establish credibility for the work, we need more mridangists to transcribe -- a scientific expt. requires establishment of credibility. The idea is that we take these different transcriptions and find the mappings. We have had some success with ML algorithms for classification of the strokes. Akshay played a set of tanis and was able transcribe his own tani using ML algos -- so it is not a lost cause!
I do not know much about mridangam -- but I heard from two colleagues who are professional (not in the concert circuit) but do accompany some musicians at IIT Madras occasionally that there are differences in schools -- but these are primarily mappings.
As long as the mapping can be done -- we can train algos to learn them. We are in fact interested in arriving at appropriate notation that experts will agree upon -- we are even willing to give different transcriptions for different schools.

-hema

[Nick H]

Wishing you success

[mahavishnu]

Hema, sounds like a very interesting project. All the best.

[Nick H]

Do keep us informed of progress, please.

[vasanthakokilam]

Hema: It sounds quite interesting indeed. If I understood right, you want the transcription of tanis to verify if your ML algorithms classify them correctly. True?

If Sri. Balaji ( forum id mridhangam, http://www.rasikas.org/forums/memberlist ... ile&u=1755) reads this thread, he can be a good candidate for you if he has time since he is quite research minded/methodology minded. You should see his matrix for how to construct a Korvai. Send him an email through the forum.

Is it a linear discriminant classifier, meaning many strokes strung together has a better chance of getting all of them right than when individually analyzed? Asking because, I think mridangists chunk several together and more or less treat it as a unit. Mridangists, please elaborate.

For example, the ending of a thani has a familiar pattern using specific strokes. The Mohara which precedes that ending climax Korvai has that "THALONGU THOM DHI, THALONGU THOM" which I hear well even though I would not recognize a Thom from Dhi individually. That is probably an LDC acting in my brain I think a good LDC algorithm would latch on to it and classify it as a Mohara and then take it apart to get at the strokes individually. Just a thought, you probably are much ahead on such matters.

See this therad where mridangamkid explains the details of what happens at the end of a thani quite well. http://rasikas.org/forums/viewtopic.php?f=8&t=7264

Best wishes for the success of your research project.

[hema]

Dear Vasanthakokilam
First, we need appropriate features and then the classifier. We have developed certain features that show a lot of promise.
We have used a variant of the spectrogram with HMMs, SVMs. If a certain sequence of strokes go together -- this is not an issue from an ML perspective. We can ensure that the algos learn the right transcription. The problem that I have is the following:
Each musician will give his/her own transcription. We need transcriptions from many musicians so that we are objective. I do have one person with me -- In a scientific experiment this is pretty inadequate -- since transcriptions are based on subjective evaluations. In standard methodology (I have done this in other contexts) -- we get different people to transcribe the same piece. We then get the correlation across various people. We remove outliers (peoples' transcriptions). Next we use the transcriptions that have at least a correlation of 0.8. These are used as ground truth. We divide the data into train and test. Perform n-fold cross validation. These are then used to transcribe new data. The errors are corrected by the experts again. Getting subjective evaluations is difficult so we settle for about 4-5 people. In contexts, where it is primarily based on giving a single rating based on listening tests -- we get 30 people to vet the results (30 because statisticians tell us that 30 iid examples are good for making a decision).
BTW I went to the link. It has a whole lot of advertisements. At the risk of giving up my identity on this forum -- if you know the person well -- can you send his mailID to my email address?
Thanks, Hema

[vasanthakokilam]

OK, Understood. I have sent you an email.

[hema]

Thanks a lot.

[shashikiran]

I know of a couple of people who are working in this area if you email [email protected] can help you connect with them

[thathwamasi]

Why not try Palghat Ramprasad? I know he holds a PhD in econometrics with specialty in mathematical and econometric modeling ? I'm not sure how much time he might have given the fact he works for a Big 3 + his teaching schedule + his own concerts and practice time. Nevertheless may be worth trying given his musical and academic background.

[hema]

Dear All
Thanks for the overwhelming response. I will try and contact most people that you have suggested.
Rgds, Hema