A Novel Source Filter Model using LSTM/Kmean Machine Learning Methods for the Synthesis of Bowed-String Musical Instruments
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. Though it may take lots of computing power in the analysis/training stage to generate all the parameters of the predictor and the granular wavetable, it is computationally efficient in the synthesis processing. Changes of pitch and dynamics can be easily achieved in real time, too. In this paper, we use the violin tones in the RWC database to show our results.
Authors: Hung-Chih Yang (National Cheng Kung University), Yiju Lin (National Cheng Kung University) and Alvin Su (National Cheng Kung University)
