A novel source filter model using LSTM/K-means machine learning methods for the synthesis of bowed-string musical instruments

Synthesis of realistic bowed-string instrument sound is a difficult task due to the diversified playing techniques and the ever-changing dynamics which cause rapidly varying characteristics. The noise part closely related to the dynamic bow-string interaction is also regarded as an indispensable part of the musical sound. Neural networks have been applied to sound synthesis for years. In this paper, a source filter synthesis model combined with a Long-Short-Term-Memory (LSTM) RNN predictor and a self-organized granular wavetable is proposed. The synthesis sound can be close to the recorded tones of a target bowed-string instrument. The timbre and the noise are both well preserved. Changes of pitch and dynamics can be easily achieved in real time, too.