Sound Source Analysis and Separation

• Main Idea Behind the Project

              Signal Processing was really one interesting course for us and we have learnt quite notable things like filtering,a bit image processing also audio processing,noise removing and some interesting applications like radar,sonar how they are designed and mechanics behind them.So finally we are assigned to prepare something of our own which has to match with those contexts and apply what we have learnt.Firstly we thought to do something on image processing but we already has some works done on those so to make something new,we moved from image to audio signal processing.That's the thinking and motivation that drove us throughout the process.

• What it does ?

              What we are trying to do is quite simple.We want to build our own system where we can give an audio source as input and it can separate the different sound component sources and generate those separated ones as separated audio files.So to build the system we used one Machine Learning Algorithm named as "Cocktail Party Algorithm".But what "Cocktail Party Algorithm" does actually ?
              It's like if you are middle of a party where noise coming from all over the sides but still how we can focus on only one person's talking - imitating the procedure over and over again.So doing so almost all the sources get separated.For human beings the same phenomenon happens which is called "Cocktail Party Effect".

• Platform we used

              The main platform we used was Python. To plot and simulate we also used MATLAB.So according to the idea we also did some modifications and at the end we got our output.But we couldn't make accurate enough due to some noise.So it is working with some additional noise.

• Difficulties we faced

              First of all we had to learn about its behind theory as without it we can't even move forward.After understanding the theory we had to think about the implementation and we chose Python because we could use some SVM(Support Vector Machine) and do easily.So we had to learn some Machine Learning Stuffs related those and some implementation only on those.So we were a bit doing-learning loop.After some days when we constructed the model we have seen that for a big enough audio file like 60 seconds or more it has some performance compromisation.So we started thinking to use hamming window to remove it some sort of.After doing so it reduced a bit but not in full.May be we have to add some additional filters and modification for that.

• Much Grateful To

              Obviously I would like to share the project credit with two of my project mates without who it would be real difficult to work correctly.So many many times grateful to -

              Shahriar Ivan

           Current Student,6th Semester,Dept. of CSE

              And

              Fahim Shahriar Shakkhor

          Current Student,6th Semester,Dept. of CSE