Title Page
Contents
Abstract 11
Chapter 1. Introduction 13
Chapter 2. Background 17
2.1. Musical Background 17
2.1.1. Historic Background 17
2.1.2. Transformation Process of Yeominlak 18
2.1.3. Chihwapyeong, Chwipunghyeong 19
2.2. Technical Background 22
2.2.1. Symbolic Music Generation 22
2.2.2. Previous works for Korean music 23
Chapter 3. Dataset 25
3.1. Yeominlak Chronological Dataset 25
3.2. Chihwapyeong, Chwipunghyeong Dataset 26
3.3. Data Statistics 29
3.3.1. Beat and Dynamics 29
3.3.2. Pitch and Duration 31
Chapter 4. Methodology 33
4.1. Data processing 33
4.1.1. Measure Alignment 33
4.1.2. Note Encoding 33
4.2. Model 35
4.2.1. Auto-regressive RNN (Proposed) 36
4.2.2. Non-autoregressive RNN 40
Chapter 5. Experiments and Evaluation 42
5.1. Training 42
5.1.1. Training and Validation Dataset 42
5.1.2. Era Condition 43
5.1.3. Training Details 44
5.2. Inference 45
5.2.1. Era-sequential Inference 45
5.2.2. Moving Window Inference 46
5.2.3. Beat Shifting 47
5.2.4. Generated Result 47
5.3. Evaluation 50
5.3.1. Metrics Evaluation 50
5.3.2. Evaluation Result 51
5.3.3. Expert Evaluation 53
Chapter 6. Discussions and Conclusion 55
Bibliography 57
국문초록 61
Table 2.1. Scores conveying Yeominlak 18
Table 3.1. Musical characteristics of each part of Yeominlak 29
Table 3.2. Part-wise pitch similarity and strong beat matched rate of Yeominlak 30
Table 5.1. Training Detailed Information 44
Table 5.2. Evaluation results for a comparison of position-conditioned (shifting position) and pitch modification techniques in an Autoregressive RNN model for mitigating overfitting 51
Table 5.3. The table comparing the evaluation results of Autoregressive and Non-autoregressive models across three above metrics 53
Figure 1.1. Diagram depicting our task. Our goal is to inferring how Chihwapyeong and Chwipunghyeong would have transformed througout ages based on the transformation of Yeominlak. 14
Figure 2.1. Yeominlak in 15 th century-Sejong Silok Score 19
Figure 2.2. Yeominlak in 19 th century-Samjukgeumbo 20
Figure 2.3. Score of Chihwapyeong 21
Figure 2.4. Score of Chwipunghyeong 22
Figure 3.1. Chihwapyeong dataset 27
Figure 3.2. Chwipunghyeong dataset 28
Figure 3.3. Yeominlak measure alignment 32
Figure 4.1. Note encoding 34
Figure 4.2. Auto-regressive RNN model figure 38
Figure 4.3. Non-autoregressive RNN model figure 41
Figure 5.1. Self-similarity matrix of 8th-era Yeominlak 43
Figure 5.2. Era-sequential inference 45
Figure 5.3. Moving window inference 46
Figure 5.4. Generated Results using Autoregressive RNN model (Chihwapyeong, 9-14 measures) 48
Figure 5.5. Generated results using non-autoregressive RNN model (Chihwapyeong, 9-14 mea- sures) 49
Figure 5.6. Decoder's attention map with two different training methods. X axis corresponds to the decoder time step (note) and Y axis corresponds to the encoder time step (note). 52