Automatic Synthesis Technology of Music Teaching Melodies Based on Recurrent Neural Network

Table 4

Prediction errors of acoustic features of different model configurations on the test set.

Model	Number of layers	F0 correlation	F0 RMSE	LE	LCD
DCNN	1	0.73	41.39	3.95	6.25
	2	0.85	40.72	2.67	5.12
	3	0.89	38.85	2.53	4.87
	4	0.84	39.03	2.51	4.82
	5	0.82	39.73	2.54	4.83
LSTM	1	0.86	39.46	3.76	5.14
	2	0.85	37.12	2.64	4.69
	3	0.84	36.80	2.68	4.23
	4	0.89	36.71	2.63	4.52
	5	0.88	38.62	3.85	4.34
Our model	1	0.86	38.94	2.67	5.09
	2	0.85	35.23	2.59	4.72
	3	0.89	35.89	2.52	4.24
	4	0.88	35.42	2.63	4.12
	5	0.87	35.73	2.51	4.35