Neural Networks And Ensemble Based Architectures To Automatic Musical Harmonization: A Performance Comparison

Figures & data

Figure 1. Block diagram showing the proposed methodology, step by step, described in order.

Figure 2. Percentage of notes, major and minor chords present in 20% of the database, for each of the 12 possible notes considered in Western Tonal Music.

Table 1. Average loss and accuracy, $F_{1M}$, MCC and $\kappa$ percentage for each model tested. Neurons are shown in parentheses.

Experiment	Loss	ACC (%)	F1%)	MCC (%)	$\kappa$ (%)
MLP1 (256)	1.7356	46.89	15.69	33.19	32.13
MLP2 (64)	1.7376	46.99	14.04	33.20	32.06
RBF (256)	1.7699	46.33	10.17	31.97	30.61
ELM (256)	2.1070	47.06	16.37	33.21	31.90
ESN (256)	2.1931	44.64	9.31	29.31	27.31
LSTM (64)	1.8485	45.65	15.59	32.12	31.02
ENS6	2.0076	46.94	15.55	33.08	31.56
ENS3	1.7743	47.02	14.04	33.05	31.63
ENS2	1.7676	46.75	14.19	32.71	31.29

Table 2. Overall ranking using Borda count method considering different metrics and algorithm results.

Experiment	Loss	F1	MCC	$\kappa$	Total
MLP1	9	8	7	9	33
MLP2	8	4	8	8	28
ELM	2	9	9	7	27
ENS2	7	5	4	4	20
ENS6	3	6	6	5	20
ENS3	5	3	5	6	19
LSTM	4	7	3	3	17
RBF	6	2	2	2	12
ESN	1	1	1	1	4

Figure 3. Melody of the song America represented with (a) score notation and (b) original and generated harmony by each evaluated model.

Figure 4. Normalized confusion matrix for classifying chords using the MLP1 model.