Digital Signal Processing Reference
In-Depth Information
Basic elements for the drum beats are hi-hat, snare and bass. The combina-
tion of those elements creates various rhythms such as bossanova, shufle, dub,
and drum'n bass. The automatic generation of the drum beats in this paper is
essentially reduced to determining the states of 16 sixteenth notes respectively.
We have 16 time steps where the state must be determined. Each state can be
'Touched' or a mute sound.
In order to determine one drum beat, 16 sixteenth notes were generated for
each element (i.e, hi-hat, snare and bass). The beat determination is performed
for each adjacent two notes. Therefore, the 16 notes are grouped as 8 states
variables. Each state can have one of four possible values such as 00, 01, 10 and
11. We grouped the adjacent beats to control the probability each possible values
so that the generated beats are more reasonable and plausible. For example, the
hi-hat beats should have more frequent beats than snare or bass. However, snare
or bass should have lower probability for the states such as 11, 01 and 10. Fig.
3 shows how the shue rhythm can be generated with our approach.
Fig. 4.
Composition of shufle rhythm
3.3 Automated Classification of Generated Music
It is hard to quantify the evaluation of music because the understanding or
emotional acceptance of music pieces is extremely subjective process. Therefore,
it might be almost impossible to design a well-defined functions or measurement
for the evaluation of music. Therefore, we employed the machine learning based
evaluation system based on backpropagation neural network.
The backpropagation is an example of supervised learning system. According
to the delta learning rule, the network adjusts the weights to obtain optimized
connections to recognize the data set of which recoginition result is given. After
the learning phase, the system can be utilized to recognize any data set based
on the previous learning. Therefore, the method is proper to be quantify the
subjective evaluation of music based on the user tastes, and can be utilized for
the evaluation of various music pieces.
The machine learning of the backpropagation method is based on the input
vector and weights of the links that connects the input to the output vector.
