Model-based synthesis of plucked string instruments by using a class of scattering recurrent networks

Area of Science:

• Acoustics
• Music Synthesis
• Machine Learning

Background:

• Traditional electronic music synthesis methods like frequency modulation and wavetable have limitations in realistically modeling complex acoustic phenomena.
• Accurate physical modeling of instruments is crucial for creating authentic synthesized sounds.

Purpose of the Study:

• To propose a novel physical modeling method for synthesizing plucked-string tones using recurrent neural networks.
• To develop a virtual model capable of mimicking the dynamic behavior of acoustic strings.

Main Methods:

• A scattering recurrent network (SRN) was developed based on the physics of acoustic strings.
• The SRN was trained using measured vibration data from a plucked string through supervised learning.

Main Results:

• The trained SRN accurately models string dynamics, generating tones very close to acoustic counterparts.
• The "virtual string" responds realistically to different plucking actions, surpassing traditional synthesis techniques.

Conclusions:

• The proposed SRN-based physical modeling offers a powerful new technique for electronic music synthesis.
• The method demonstrates encouraging results in modeling and synthesizing plucked-string sounds, as shown with a cello "A"-string simulation.