http://orcid.org/0000-0002-5870-2609
Figures & data
Figure 1. Piano keyboard (Tornblom, Citation2008).
Figure 2. Synthesia screenshot.
Figure 3. MIDI messages.
Figure 4. LSTM memory cell (Donges, Citation2018).
Figure 5. Three-layer LSTM network (Graves, Citation2014).
Figure 6. Proposed architecture.
Figure 7. Song format conversion process.
Figure 8. Encoding method.
Figure 9. Transposition from C major to F major (Clementstheory.com, Citation2018).
Figure 10. Augmentation process.
Figure 11. Sliding window.
Figure 12. NN training.
Figure 13. Music generating method.
Figure 14. Dataset key count. X-axis: Key in numbers, y-axis: Normalized frequency.
Figure 15. Note occurrences. X-axis: Piano key in numbers, y-axis: Un-normalized frequency.
Figure 16. Architecture of the NN.
Figure 17. Testing LSTM layer sizes. (a) Training F-measure, (b) training loss, (c) testing F-measure, (d) testing loss.
Figure 18. Testing dropout values. (a) Training F-measure, (b) training loss, (c) testing F-measure, (d) testing loss.
Figure 19. Testing different optimizers. (a) Training F-measure, (b) training loss, (c) testing F-Measure, (d) testing loss.
Figure 20. Testing augmentation values. (a) Training F-measure, (b) training loss, (c) testing F-measure, (d) testing Loss.
Figure 21. Testing batch sizes. (a) Training F-measure, (b) training Loss, (c) testing F-measure, (d) testing loss batch sizes 128 (red), 256 (blue), 512 (orange), and 1024 (grey).
Figure 22. Testing window sizes. (a) Training F-measure, (b) training loss, (c) testing F-measure, (d) testing loss.
