Pith. sign in

REVIEW 3 major objections 5 minor 24 references

Apollo: An Interactive Environment for Generating Symbolic Musical Phrases using Corpus-based Style Imitation

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Apollo turns a user-supplied MIDI corpus into a trainable model that generates new musical phrases in that style.

desk verdict A promising but unvalidated system description of an interactive corpus-based style imitation environment; the integration is new, but Model1's objective is unspecified and no evaluation is provided. read the letter →

arxiv 2504.14055 v1 pith:MEPW6F4O submitted 2025-04-18 cs.HC cs.LGcs.SD

classification cs.HCcs.LGcs.SD
keywords interactivemachinelearningstyleimitationgenerativemusicMIDIcorpuscomputer-assistedcompositionmusicalmetacreationconstraint-basedgenerationVAE
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Apollo is an interactive music-generation environment whose central claim is that artists and researchers can generate new symbolic musical phrases in the style of a corpus they choose, without writing code. The paper argues that existing generative systems are mostly pre-trained black boxes, and that opening up corpus management, model selection, training, and generation-parameter control addresses a real gap. The system hosts two models, a constraint-based algorithm developed for Apollo and the MusicVAE, and exposes their parameters through automatically rendered GUI controls. If the claim holds, Apollo provides a working alternative for hands-on style imitation and a platform for testing new models.

What carries the argument

The load-bearing mechanism is a three-phase interactive machine-learning workflow: a corpus manager for MIDI files, a model-training interface that dynamically discovers Python models dropped into the system and renders their parameter controls from a JSON specification, and a generation interface with the same automatic control rendering. Inside the hosted constraint-based model, the key algorithmic choice is generating rhythm before pitches, guided by statistical features extracted from the corpus, with user preferences reflected in parameters such as melodic typicality and note density. This two-step strategy keeps the search manageable and gives a short training-to-generation feedback loop, while MusicVAE supplies a latent-vector-based alternative route.

What would settle it

A listening test in which participants are asked whether a generated phrase comes from the style of corpus A or corpus B: if listeners cannot distinguish phrases generated from two distinct corpora at above chance, the claim that the system imitates the corpus style fails.

Watch

Extended reading notes

Core claim

The paper presents Apollo as a system that makes corpus-based style imitation interactive throughout the machine-learning loop. Its contribution is the workflow: users assemble and edit a MIDI corpus, choose and train a style-imitation model through a GUI, then generate new phrases whose melodic typicality, harmonic following, number of measures, and note density they can adjust. For the in-house model, generation is done by a constraint solver that first builds rhythm and then selects pitches, a necessary division because the joint search space was too large; for MusicVAE, new 16-bar sequences come from sampling latent vectors associated with the user's corpus. The authors claim this design lets musicians and researchers experiment with stylistic imitation and quickly test new models.

Load-bearing premise

The system's style-imitation claim rests on the assumption that the patterns it extracts from the uploaded music files, together with generating rhythm before pitches, capture enough of the corpus style that listeners would attribute the output to that style and not to generic generation.

Editorial extensions

If this is right

  • Users can train a style-imitation model on their own MIDI corpus and generate new phrases without programming, because training and generation are exposed through the GUI.
  • New Python models can be added to the system by placing them in the model discovery folder with a JSON parameter file, and the GUI will generate controls for them automatically.
  • Generation parameters such as melodic typicality, harmonic following, number of measures, and note density give the user direct influence over the character of the output from the trained model.
  • Generated phrases can be exported as MIDI or streamed to an external sequencer, making the system usable as a seed-material tool inside a musician's existing workflow.
  • Because Model1 separates rhythm and pitch generation to keep the search space tractable, the quality of its style imitation depends on the statistical features used for each step.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper would be strengthened by a perceptual test, and the natural experiment is to generate phrases from two stylistically distinct corpora and ask listeners whether they can tell which corpus produced each phrase; the current paper reports no such evaluation.
  • The same GUI-plus-model-discovery architecture could serve as an evaluation harness: identical corpora could be run through Model1 and MusicVAE and compared on style fidelity, a comparison the paper does not make.
  • The rhythm-first design implies a testable hypothesis, namely that style fidelity in this system depends more heavily on rhythmic statistics than on pitch statistics, since pitches are selected after rhythm is fixed.
  • The ability to manage multiple corpora points toward style mixing, which the paper mentions as a possibility but does not implement or evaluate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Apollo, an interactive desktop/web environment for corpus-based symbolic music generation. Users can upload and manage MIDI corpora, train a generative model (either a custom constraint-based model called Model1 or MusicVAE), and generate trio-style musical phrases by adjusting GUI parameters such as melodic typicality, harmonic following, number of measures, and note density. The system architecture, implementation stack (MEAN with Electron), and the three-phase interactive workflow are described in detail. The paper claims that Apollo enables artists and researchers to generate new musical phrases in the style of a user-supplied corpus, but it provides no evaluation, no output examples, and no specification of Model1's statistical features or optimization objective.

Significance. If the central claims are substantiated, Apollo would be a valuable contribution to interactive machine learning for music, since it gives users open control over corpus preparation, model selection, and generation parameters, and it supports dynamic discovery of new Python models. The system design is coherent and the implementation details are sufficiently concrete to be reproduced. The strongest strengths are the explicit focus on the interactive workflow and the integration of existing MusicVAE machinery with a custom constraint-based model. However, the manuscript currently does not provide evidence that the generated phrases are stylistically imitative, which is the core promise of the title and abstract; the principal model is underspecified and no empirical validation is reported. As a system-description paper the contribution is plausible, but the overclaim without supporting evidence is a significant gap.

major comments (3)
  1. [Machine Training for Style Imitation] Model1's statistical features and constraint-solver objective are never enumerated. The text states that the solver 'tries to find an optimal solution that exhibits stylistic characteristics of the corpus,' but it does not define those characteristics or the objective function. Given that the Background section defines style imitation as generation that an unbiased observer would label as belonging to the corpus style, the paper currently provides no way to audit whether the optimizer's notion of style matches that definition. Please specify the extracted features, the optimization objective, and the user-controlled constraints, or cite a prior publication that defines them.
  2. [Machine Training for Style Imitation] The two-step rhythm-then-pitch generation is justified solely by the claim that the joint search space was too large, but the paper gives no evidence that this decomposition preserves the joint rhythmic and pitch statistics that characterize a corpus style. This is load-bearing because the style-imitation claim depends on the generated output retaining the corpus-specific stylistic character. Please provide an argument, a reference, or empirical evidence (e.g., comparing output statistics with corpus statistics) that the two-step generation does not destroy stylistic fidelity.
  3. [Discussion and Future Work] The paper provides no evaluation in any form: no generated output examples (notation, MIDI, or audio), no listening tests, and no objective style metrics. The Future Work section explicitly states that user evaluation is planned, which is appropriate for a system paper, but the abstract's claim that Apollo 'makes it possible' to generate style-imitative phrases is not supported as written. Please add at least a minimal demonstration of generated output and some form of style evaluation, or explicitly reposition the paper as a description of an unvalidated prototype.
minor comments (5)
  1. [Throughout] The model name is written inconsistently as 'MusicV AE' and 'MusicVAE'; please use a single consistent spelling.
  2. [Implementation] Please spell 'MongoDB' consistently, and note that 'CSS4' is not a standard version name; 'CSS3' or a more specific reference would be clearer.
  3. [Discussion] There are several grammatical and typographical errors, for example 'The opportunity to manage more than one corpus enables for easy ways to create new corpora' and 'this support can enable an environment'; these should be corrected.
  4. [Generation Parameters] The JSON example is not formatted as a proper code listing; please present it in a monospaced block so that it is readable and unambiguous.
  5. [Figures] The figure captions are very terse; adding one or two sentences to each figure explaining what the user sees and how it supports the workflow would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Apollo is a system description whose style-imitation claim is under-specified but does not reduce by construction to a fitted parameter or self-citation.

full rationale

The paper reports no quantitative prediction, no fitted parameter that is later renamed as an evaluation, and no uniqueness theorem that forces a choice. Model1 is described qualitatively: it 'extracts statistical features from the MIDI files in the training corpus' and 'the constraint solver tries to find an optimal solution that exhibits stylistic characteristics of the corpus' (Machine Training for Style Imitation). This is the intended corpus-based training/generation mechanism, not a circular derivation: the text provides no equation, feature list, or evaluation metric that equates the output's quality with the solver's own objective. The MusicVAE path similarly samples latent vectors recovered from the corpus, which is a standard generative procedure rather than a disguised prediction. Several citations are to work by the same group (Pasquier et al. 2016; Anderson, Eigenfeldt, and Pasquier 2013; Eigenfeldt and Pasquier 2013), but these are contextual definitions and related-work descriptions, not load-bearing justifications for the paper's central claim. The absence of an external listener study is a real limitation, and the authors themselves defer user evaluation to future work, but that is a completeness concern, not circularity. No circular step can be quoted and exhibited from the text, so the score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

Apollo wraps two existing generative approaches in an interactive shell. The central capability claim rests on untested domain assumptions about how well MIDI-derived statistics and the chosen models preserve style, and on fixed structural choices that shape all outputs. No new conceptual entities are introduced.

free parameters (2)
  • MusicVAE sequence length = 16 bars
    The MusicVAE workflow splits and generates fixed 16-bar sequences; this hand-chosen window bounds every generated phrase and is not derived from the corpus.
  • Model1 output part layout = 3 parts (bass, melody, drums)
    Model1 always generates a trio of bass, melody, and drum parts; the fixed output structure is imposed on every corpus and does not adapt to the corpus content.
assumptions (4)
  • domain assumption Style can be captured by statistical features extracted from MIDI files and reproduced by a constraint solver.
    Model1's training, described in 'Machine Training for Style Imitation', relies on this without validation; no evidence links the extracted features to listener judgments of style.
  • ad hoc to paper Generating rhythm before pitches preserves enough stylistic coherence for style imitation.
    The paper says this decomposition was chosen out of necessity because the joint search space was too large; the stylistic cost of the decomposition is unmeasured.
  • domain assumption MusicVAE trained on the user corpus produces output that counts as style imitation.
    The paper adopts MusicVAE (Roberts et al. 2018) and samples latent vectors from the user corpus, but reports no evaluation of style fidelity in Apollo's use case.
  • domain assumption Interactive machine learning improves the creative value of generative music systems.
    This motivating premise is cited to IML literature (Fails and Olsen Jr 2003) but not tested here; the authors state user evaluation is future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Apollo: An Interactive Environment for Generating Symbolic Musical Phrases using Corpus-based Style Imitation." pith.science (2026). https://pith.science/paper/MEPW6F4O

@misc{pith2026250414055,
  author       = {Pith},
  title        = {Pith review of: Apollo: An Interactive Environment for Generating Symbolic Musical Phrases using Corpus-based Style Imitation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEPW6F4O}},
  note         = {Machine review of arXiv:2504.14055}
}
read the original abstract

With the recent developments in machine intelligence and web technologies, new generative music systems are being explored for assisted composition using machine learning techniques on the web. Such systems are built for various tasks such as melodic, harmonic or rhythm generation, music interpolation, continuation and style imitation. In this paper, we introduce Apollo, an interactive music application for generating symbolic phrases of conventional western music using corpus-based style imitation techniques. In addition to enabling the construction and management of symbolic musical corpora, the system makes it possible for music artists and researchers to generate new musical phrases in the style of the proposed corpus. The system is available as a desktop application. The generated symbolic music materials, encoded in the MIDI format, can be exported or streamed for various purposes including using them as seed material for musical projects. We present the system design, implementation details, discuss and conclude with future work for the system.

Figures

Figures reproduced from arXiv: 2504.14055 by the authors.

Figure 1
Figure 1. Browsing Interface MIDI files in the training corpus. The musical content is gen￾erated in two steps, as the rhythm structure is generated be￾fore specific pitches are selected. This decision was made out of necessity, as the search space was simply too large when rhythm and pitch were generated concurrently. For both steps, the constraint solver tries to find an optimal solu￾tion that exhibits stylistic characteris… view at source ↗
Figure 4
Figure 4. Interactive Workflow • Melodic Typicality: This parameter controls the degree to which the generated melody adheres to the style learned by the trained model. • Harmonic Following: The extent to which harmonies in the generated musical phrase follows each other coher￾ently. • Number of measures: The number of measures each gen￾erated musical phrase contains. • Note Density: The average amount of notes found per meas… view at source ↗
Figure 3
Figure 3. Generative Interface the definition of those parameters. The specification model parameters is composed of a list of attributes for each pa￾rameters: a default value, description, display name (for the GUI), a maximum value, a minimum value, the program￾matic name and the parameter type. An example from the MusicVAE model is: { "default": 0.001, "desc": "amount of noise added to latent vector", "display_name": "Nois… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: System Architecture Discussion Developing the Apollo system poses many interesting chal￾lenges such as manipulating a large set of binary files (for the MIDI music corpus) and transmitting them over the inter￾net connection with limitations of HyperText Transfer Pro￾to…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    Agostini, A., and Ghisi, D. 2013. Real-time computer-aided composition with bach. Contemporary Music Review 32(1):41--48

  3. [3]

    Anderson, C.; Eigenfeldt, A.; and Pasquier, P. 2013. The generative electronic dance music algorithmic system (gedmas). In In Proceedings of the Second International Workshop on Musical Metacreation (MUME 2013) 2013

  4. [4]

    Assayag, G.; Bejerano, G.; Dubnov, S.; and Lartillot, O. 2001. Automatic modeling of musical style. In 8 \`e mes Journ \'e es d'Informatique Musicale , 113--119

  5. [5]

    Assayag, G. 1998. Computer assisted composition today. In First Symposium on Music and Computers . Corfu Greece: Ionian University

  6. [6]

    Briot, J.-P.; Hadjeres, G.; and Pachet, F. 2017. Deep learning techniques for music generation-a survey. arXiv preprint arXiv:1709.01620

  7. [7]

    Domingos, P. 2012. A few useful things to know about machine learning. Communications of the ACM 55(10):78--87

  8. [8]

    Dubnov, S.; Assayag, G.; Lartillot, O.; and Bejerano, G. 2003. Using machine-learning methods for musical style modeling. Computer 36(10):73--80

Show all 24 references
  1. [9]

    Eigenfeldt, A., and Pasquier, P. 2013. Evolving structures for electronic dance music. In Proceedings of the 15th Annual Conference on Genetic and Evolutionary Computation , 319--326. ACM

  2. [10]

    A., and Olsen Jr, D

    Fails, J. A., and Olsen Jr, D. R. 2003. Interactive machine learning. In Proceedings of the 8th International Conference on Intelligent User Interfaces , 39--45. ACM

  3. [11]

    D., and Vico, F

    Fern \'a ndez, J. D., and Vico, F. 2013. Ai methods in algorithmic composition: A comprehensive survey. Journal of Artificial Intelligence Research 48:513--582

  4. [12]

    Fiebrink, R., and Caramiaux, B. 2016. The machine learning algorithm as creative musical tool. CoRR abs/1611.00379

  5. [13]

    Hadjeres, G.; Pachet, F.; and Nielsen, F. 2017. Deepbach: a steerable model for bach chorales generation. In Proceedings of the 34th International Conference on Machine Learning - Volume 70 , ICML'17, 1362--1371. JMLR. org

  6. [14]

    Herremans, D.; Chuan, C.-H.; and Chew, E. 2017. A functional taxonomy of music generation systems. ACM Computing Surveys (CSUR) 50(5):69

  7. [15]

    Liang, F. 2016. Bachbot: Automatic composition in the style of bach chorales. University of Cambridge 8:19--48

  8. [16]

    Maxwell, J. B. 2014. Generative music, cognitive modelling, and computer-assisted composition in musicog and manuscore . Ph.D. Dissertation, ? by Home Dept & Faculty of Senior Supervisor: Special Arrangements

  9. [17]

    Papadopoulos, G., and Wiggins, G. 1999. AI methods for algorithmic composition: A survey, a critical view and future prospects. In AISB Symposium on Musical Creativity , volume 124, 110--117. Edinburgh, UK

  10. [18]

    Pasquier, P.; Eigenfeldt, A.; Bown, O.; and Dubnov, S. 2016. An introduction to musical metacreation. Computers in Entertainment (CIE) 14(2):2

  11. [19]

    J.; Fels, S.; et al

    Poupyrev, I.; Lyons, M. J.; Fels, S.; et al. 2001. New interfaces for musical expression. In CHI'01 Extended Abstracts on Human Factors in Computing Systems , 491--492

  12. [20]

    Roberts, A.; Engel, J.; Raffel, C.; Hawthorne, C.; and Eck, D. 2018. A hierarchical latent vector model for learning long-term structure in music. arXiv preprint arXiv:1803.05428

  13. [21]

    L.; Santos, J

    Sturm, B. L.; Santos, J. F.; Ben-Tal, O.; and Korshunova, I. 2016. Music transcription modelling and composition using deep learning. arXiv preprint arXiv:1604.08723

  14. [22]

    Thio, V.; Liu, H.-M.; Yeh, Y.-C.; and Yang, Y.-H. 2019. A minimal template for interactive web-based demonstrations of musical machine learning. arXiv preprint arXiv:1902.03722

  15. [23]

    Waite, E.; Eck, D.; Roberts, A.; and Abolafia, D. 2016. Project magenta. https://magenta.tensorflow.org

  16. [24]

    Wright, M.; Freed, A.; et al. 1997. Open soundcontrol: A new protocol for communicating with sound synthesizers. In ICMC

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.