TY - GEN
T1 - Synthesis of wind instruments and their blowing noise using a lstm time varying source filter model
AU - Chen, Ju Yen
AU - Yang, Hung Chih
AU - Su, Wen Yu
N1 - Funding Information:
It is noted that the prediction error of the current LSTM predictor is still high. Therefore, the sound quality is compromised. The efficiency is reduced because the granular wavetable databases in the system memory proposed in [12] has to be accessed frequently. It is This work is supported by MOST Taiwan in part under contract 107-2221-E-006-197-MY3.
Publisher Copyright:
© 2021 150th Audio Engineering Society Convention, AES 2021. All rights reserved.
PY - 2021
Y1 - 2021
N2 - Digital Waveguide Filters have been applied to synthesis of wind instruments for years. To design the filter coefficients to synthesize the timbre of a particular instrument usually takes lots of time and effort. A source filter model combined with a long short-term memory(LSTM) recurrent neural network for the synthesis of violin has been successful in this respect, but it lacks the synthesis of the noise part that is important for any playing wind instruments. In this paper, we adopt the noise synthesis method of DWF and apply it to the LSTM based source filter model to take the advantages of both methods. The French horn tones recorded in the Real World Computing(RWC) database are used to demonstrate the work. The synthesis tone sounds close to the original tone. Sound files are provided. The proposed method can be efficiently implemented as a VST plugin.
AB - Digital Waveguide Filters have been applied to synthesis of wind instruments for years. To design the filter coefficients to synthesize the timbre of a particular instrument usually takes lots of time and effort. A source filter model combined with a long short-term memory(LSTM) recurrent neural network for the synthesis of violin has been successful in this respect, but it lacks the synthesis of the noise part that is important for any playing wind instruments. In this paper, we adopt the noise synthesis method of DWF and apply it to the LSTM based source filter model to take the advantages of both methods. The French horn tones recorded in the Real World Computing(RWC) database are used to demonstrate the work. The synthesis tone sounds close to the original tone. Sound files are provided. The proposed method can be efficiently implemented as a VST plugin.
UR - http://www.scopus.com/inward/record.url?scp=85113873985&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85113873985&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:85113873985
T3 - 150th Audio Engineering Society Convention, AES 2021
BT - 150th Audio Engineering Society Convention, AES 2021
PB - Audio Engineering Society
T2 - 150th Audio Engineering Society Convention, AES 2021
Y2 - 25 May 2021 through 28 May 2021
ER -