Dull and boring part of Research..

As everyone thinks research is very interesting, innovative and tough, it is sad that sometimes we tend to do something which is already done with slight changes, complexify it and publish and get publications. And as we say it, we tend to do it ourselves. Like taking an already attacked application, already used methods do some permutation of different methods and think we are doing great research. May it will be much better if we try to think afresh about a new, unattacked  application and devise a simple method to solve it, even if it is very simple. Given that we do lot of literature survey, we get many new ideas to use them, but we should not get carried away by what others have already done. We may get inspired from others work, but for our own research we must have our own different, novel ideas. As it is always not possible that our new ideas work, we tend to do this dull and boring part of research. Hope we will get over it and make it more interesting.

Effect of high frequencies in audio signals and intricacies in audio research

A nice article describes why high sampling rates are not required for audio. The audio signal is the most intriguing of all signals. It is quasi-periodic but non-stationary over long  periods above 20 ms.

I have been evaluating instants of maximum excitations or epochs and I was using 3 Core i5 systems simultaneously as I had to the run the same code using different parameters to get the best performance and accuracy. I have re-refined my output recursively so as to get the best accuracy and identification rates.

On the random way to research

My another blog describes my early stages of my research. I am well into my 3rd year of my PhD, having redefined, refined, re-refined my problem statement. Having gone through a lot of research papers, journals related to my problem, some very much theoretical, application oriented and some seemly useless papers.

The problem regarding my problem statement is the solution is a random entity. The problem is seemingly NP-hard. The real problem is it is difficult and challenging. The problem statement is: Given a single channel audio signal containing mixture of only two sources: 1) speech, 2) non-speech, the output should be two separated channels one containing only speech and other non-speech signal. Now, its a supervised learning where dictionaries for both speech and non-speech need to be learnt before testing on a mixed audio.

Now, I have seen lot of good papers on sparse dictionary learning, having finally formulated the problem which is quite difficult and has lot of applications if I solve or solve to some extent. Most of the past dictionary learning has been used for object tracking in videos, image classification, few for speech recognition  and denoising but very few for source separation.

Having done the literature survey, where most of the good papers are published   in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Transactions on Signal Processing,  International Conference on Machine Learning, IEEE Workshop on Machine Learning for Signal Processing and Journal of Machine Learning Research. 

I see source separation of guitar music and other sources as one aspect of my problem. A recording of a guitar tune played by me: