Universitas Hamzanwadi, Indonesia
BibTex Citation Data :
@article{JMASIF78593, author = {Imam Fathurrahman and Muhammad Djamaluddin and Ramli Ahmad and Hariman Bahtiar and Muhamad Sadali and Zulkipli Mansyur}, title = {Effects of Temperature Sampling and Tempo Variation on LSTM-Based Music Generation}, journal = {Jurnal Masyarakat Informatika}, volume = {17}, number = {2}, year = {2026}, keywords = {Neural music generation, Long Short-Term Memory, temperature sampling, dynamic tempo, MIDI, musical diversity}, abstract = { Long Short-Term Memory (LSTM) models for symbolic music generation may produce repetitive sequences under deterministic decoding. This study examines how sampling temperature and externally imposed tempo variation affect the statistical characteristics of generated MIDI music. A factorial experiment combined four decoding settings (Argmax and temperature sampling at T = 0.6, 0.7, and 1.0) with fixed- and dynamic-tempo rendering. Using one trained LSTM checkpoint and controlled seeds, 40 MIDI files were generated across eight conditions, with five runs per condition. The dataset was split by MIDI file before sequence construction to reduce data leakage. Generated outputs were assessed using pitch entropy, repetition ratios, N-gram overlap, pitch-class distribution similarity, rhythmic diversity, and tempo statistics. At T = 1.0, mean pitch entropy was highest (5.57 ± 0.13 bits), immediate repetition lowest (0.10 ± 0.01), and Chroma Jensen-Shannon Divergence lowest (0.0244 ± 0.0163). Higher temperatures also reduced trigram overlap with the validation corpus, indicating a diversity-similarity trade-off. Dynamic-tempo rendering introduced timing variation without changing pitch-based metrics under matched sampling conditions. These results distinguish effects of token sampling from tempo rendering but do not establish perceptual musical quality. }, issn = {2777-0648}, pages = {229--249} doi = {10.14710/jmasif.17.2.78593}, url = {https://ejournal.undip.ac.id/index.php/jmasif/article/view/78593} }
Refworks Citation Data :
Long Short-Term Memory (LSTM) models for symbolic music generation may produce repetitive sequences under deterministic decoding. This study examines how sampling temperature and externally imposed tempo variation affect the statistical characteristics of generated MIDI music. A factorial experiment combined four decoding settings (Argmax and temperature sampling at T = 0.6, 0.7, and 1.0) with fixed- and dynamic-tempo rendering. Using one trained LSTM checkpoint and controlled seeds, 40 MIDI files were generated across eight conditions, with five runs per condition. The dataset was split by MIDI file before sequence construction to reduce data leakage. Generated outputs were assessed using pitch entropy, repetition ratios, N-gram overlap, pitch-class distribution similarity, rhythmic diversity, and tempo statistics. At T = 1.0, mean pitch entropy was highest (5.57 ± 0.13 bits), immediate repetition lowest (0.10 ± 0.01), and Chroma Jensen-Shannon Divergence lowest (0.0244 ± 0.0163). Higher temperatures also reduced trigram overlap with the validation corpus, indicating a diversity-similarity trade-off. Dynamic-tempo rendering introduced timing variation without changing pitch-based metrics under matched sampling conditions. These results distinguish effects of token sampling from tempo rendering but do not establish perceptual musical quality.
Article Metrics:
Last update:
Last update: 2026-11-19 13:57:10
The authors who submit the manuscript must understand that the article's copyright belongs to the author(s) if accepted for publication. However, the author(s) grant the journal right of first publication with the work simultaneously licensed under a Creative Commons Attribution-ShareAlike 4.0 International License. Authors should also understand that their article (and any additional files, including data sets, and analysis/computation data) will become publicly available once published under that license. By submitting the manuscript to Jmasif, the author(s) agree with this policy. No special document approval is required.
The author(s) guarantee that:
The author(s) retain all rights to the published work, such as (but not limited to) the following rights:
Suppose the article was prepared jointly by more than one author. Each author submitting the manuscript warrants that all co-authors have given their permission to agree to copyright and license notices (agreements) on their behalf and notify co-authors of the terms of this policy. Jmasif will not be held responsible for anything arising because of the writer's internal dispute. Jmasif will only communicate with correspondence authors.
Authors should also understand that their articles (and any additional files, including data sets and analysis/computation data) will become publicly available once published. The license of published articles (and additional data) will be governed by a Creative Commons Attribution-ShareAlike 4.0 International License. Jmasif allows users to copy, distribute, display and perform work under license. Users need to attribute the author(s) and Jmasif to distribute works in journals and other publication media. Unless otherwise stated, the author(s) is a public entity as soon as the article is published.