2.2. Solving the problem

2.2.1. Use of Spectral Analysis

The ear hears all sound as a composition of individual discrete frequencies, each frequency being represented by a sine wave tone of varying magnitude and phase. (The sine wave is the simplest waveform containing no harmonics). As drum sounds are characterised by their tonal harmonics it would seem wise to perform some sort of spectral analysis on each sound before attempting recognition.

2.2.2. Pre-processing

Assuming the method of recognition will use some sort of spectral analysis as its input, then providing a frequency spectrum that is as clear as possible is essential. The overspill of drum sounds from adjacent beats, combined with background noise will cause losses in clarity, especially in quieter drum hits. A method for reducing background noise must therefore be devised.
The amplitude of each hit within a break compared with the other hits of similar type will determine whether the drum can be classed as a hard or light hit.

2.2.3. Neural Networks as a method of recognition

Neural networks are largely used in recognition systems where the problem is generally ill-defined. This is usually because the data to be recognised does not conform to a specific set of rules enabling recognition algorithms to easily be devised (recognition algorithms being a requirement in AI methods).
Neural networks also model parallel processing which greatly speeds up the recognition process, and in general have less computational requirements when running. A net can also easily be adapted or retrained to accommodate new data.
It was decided to use Neural Networks in this project due to the following reasons:

  • Large elements of noise will be present in some of the subject data

  • Data patterns of a specific type will be very varied

  • The recognition method used in the proposed application will need to be fast (in order to avoid hindrance to the user)

The disadvantage of neural nets is the slight unpredictability of the training methods, however experimenting with different training parameters is straight forward.
Thus a number of different types of neural network will be analysed in order to see how each network compares with each other.
In addition a certain degree of post-processing will be looked at in order to make further classification distinctions based on more general characteristics of the drum sounds.