SOUND VARIATIONS BY REECURRENJT NEURAL NETWORLK
S'YNITHESIS
Ken'ichi Ohya
Dept. of Electronics and Computer Science,
Nagano National College of Technology,
716 Tokuma, Nagano City, Nagano 381-8550, JAPAN
E-mail: ohya~gei.nagano-nct. ac. jp
ABSTRACT: A new sound synthesis techniqlue, recurrent neural network sound synthesis was shown
in 1995. Recurrent neural network is a neural network that each neuron connects recurrentlyr.
In this paper, sound variations by this recurrent neural network model are shown, One variation is from
relatively small network, and another variation is from relatively large network. The former is suitable for
realtime sound synthesis, and the latter is for modifying the sound after resynthesis of the original sound.
1 Introduction
Study of sound synthesis has a long history in computer music and many models have been presented.
A new sound synthesis model using recurrent neural network was shown in 1995, and its capability of
learning nonlinear dynamics was presented [Ohya, 1995J.
In this paper, sound variations using recurrent neural network model are shown. One sound variation is
from relatively small network, and another sound variation is from large network.
Recurrent neural network is a neural network that each neuron connects recurrently. RNNS (Recurrent
Neural Network Synthesis) belongs to nonlinear sound synthesis. The following is some features of RNNS:
1) dynamics of neurons are directly used for waveform itself 2) RNNS is like FM synthesis, each neuron
corresponds to each operator in FM synthesis model 3) complex waveforms are produced from relatively
small number of neurons 4) resynthesis is also possible with relatively large number of neurons by learning
algorithm.
2 Single Neuron Model and Recurrent Neural Network
As a single neuron model, a continuous-time, continuous-variable neuron model is adopted. Therefore output
value from any single neuron can be directly used for waveform.
Equation of dynamics of each neuron in a recurrent neural network is given ([Ohya, 19951) as
dui
where ui(t) is the i-th unit output at a time t, 'ri a time delay constant, f(x) a sigmoid function, Ii an
external input of the i-th unit, Wq ai coneto wei;~ll *~ghtv fromr the~ j-ths nitr to the i-t unl it.V
3 Sound Variation 1: Small Network of Several Neurons
Recurrent Neural Network can generate very complex dynamics pattern because of its recurrent connections
even if it is composed of relatively small number of neurons.
As the 1st variation of RNNS, sound synthesis by small recurrent neural network is shown. This model
is useful for realtime software sound synthesis because its relatively less computation.
3.1 1 pair model
"1 pair model" is composed of one pair neuron and another output neuron (Fig. 2). Each neuron of
the pair neuron is connected each other and is also connected to itself. This pair neuron is the smallest
recurrent neural network architecture. The output neuron receives two output values from thre pair neuron
and computes output value, which is waveform itself.
0