The interests of this project lie in the area of music studio software, in particular software dealing with the processing of recorded drums in music. The problem it tackles is the recognition of drum sounds taken from a wide variety of sources. It suggests that if this can be achieved, then it is possible to produce an application that can take rhythms created by one drummer or percussionist and apply them to sounds taken from different rhythm recordings.
The method of recognition it uses is a combination of spectral analysis and neural networks. There are additional factors of noise reduction taken into consideration. Many different neural network structures are looked at and their performances evaluated, and in addition a great deal of pre-processing and data manipulating code is implemented.
The project finds that the best method of recognition uses a 60:25:5 feed forward neural network, which achieves a classification rate of 74%. This is when a static spectral representation of each drum is observed (one that does not change with time). In all cases the noise reduction algorithm has an effect of decreasing the error (by factors of around 3% - 4%) with little effect on computational overheads.
It also finds that recurrent neural networks introducing temporal representation of data, do not appear to improve the classification rate.