A new accurate model-based music synthesis technique by using recurrent neural networks
Sheng‐Fu Liang, Alvin W.Y. Su · The Journal of the Acoustical Society of America · 1998
Reproduction of the tones generated by playing a particular musical instrument electronically remains an open and difficult problem. An accurate model-based analysis/synthesis approach with respect to plucked-string instruments by using recurrent neural networks is proposed. The procedure is described as follows. A recurrent neural network corresponding to the physical model of a musical string is first configured. The vibration of a plucked string is measured and used as the training data of the neural network. The backpropagation-through-time method is used to train the network such that output of this network can be matched to the measured data as close as possible. The well-trained neural network is then used as our synthesis model for this particular string. Since the characteristics of a string keep changing over the period of vibration, it is also necessary to update the parameters of the neural network during the synthesis process. In our experiments the synthesized tones and the real tones sounded almost identical. These results will also be demonstrated during the presentation in ICA/ASA 98 if this paper is accepted. [This Work is supported by both NSF, Taiwan and Computer/Communication Lab. ITRI, Taiwan.]