Pith. sign in

REVIEW 3 major objections 4 minor 45 references

The Florence Price Art Song Dataset and Piano Accompaniment Generator

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper releases a 112-song digital catalog of Florence Price's art songs and a fine-tuned model whose accompaniments beat a baseline in blind style comparisons.

desk verdict A genuinely useful dataset outweighs a fragile listening-test claim; worth refereeing for the data, with the evaluation section needing revision. read the letter →

arxiv 2506.23130 v1 pith:3EFZC3RB submitted 2025-06-29 cs.SD eess.AS

Tao-Tao He , Martin E. Malandro , Douglas Shadle This is my paper · ORCID
classification cs.SDeess.AS
keywords FlorencePriceartsongdatasetsymbolicmusicgenerationaccompanimentfine-tuningT5transformerMIDIinfillinginformationretrieval
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

This paper releases the Florence Price Art Song Dataset: 112 of Florence B. Price's songs for solo voice and piano, manually transcribed from manuscripts and made available in MuseScore, MusicXML, MIDI, and PDF, with per-song metadata and onsets. It is presented as the most complete digital collection of Price's vocal music to date, and the first in symbolic format. The paper also fine-tunes a T5-based music infilling model on this dataset to create the Florence Price Piano Accompaniment Generator, which takes a melody and writes a piano accompaniment. In a blind listening experiment, two listeners judged accompaniments from this model as more reflective of Price's style than those from a baseline fine-tuned on nineteenth-century art songs, on both Price melodies and popular melodies. The release makes an underdocumented body of music computable and shows that a small single-composer corpus can transfer stylistic identity to a generative model.

What carries the argument

The load-bearing object is the Florence Price Art Song Dataset: 112 manually transcribed symbolic scores (MuseScore, MusicXML, MIDI, PDF, plus lyrics, metadata, and section onsets). The generative mechanism is masked multi-track MIDI infilling with a T5-style encoder-decoder transformer: every measure of the piano track is masked while the melody track is left unmasked, so fine-tuning teaches the model to fill in Price-like accompaniments. To avoid leakage, accompaniments for a Price melody are generated by a model trained on the dataset excluding that song (leave-one-out). Inference uses nucleus sampling with p=0.95 and a rhythmic temperature of 1.5, and the evaluation machinery is a pairwise blind listening test with binomial significance tests.

What would settle it

Have a second musicologist independently transcribe a random sample of the 112 songs from the same manuscript photocopies and compare note-level agreement; if substantial divergence appears, the dataset's fidelity—and therefore the model's training signal—is in doubt.

Watch

Extended reading notes

Core claim

The discovery, stated in the paper's own terms, is that the released 112-song dataset is the most complete digital collection of Florence Price's vocal works and the first symbolic-format repository of her music, and that the Florence Price Piano Accompaniment Generator, obtained by fine-tuning a T5-based multi-track MIDI infilling model on that dataset, generates accompaniments that listeners perceive as reflective of Price's style significantly more often than a baseline fine-tuned on a large nineteenth-century art-song corpus. Table 1 reports 82/137 pairwise wins on Price melodies (p=0.026) and 99/137 on popular melodies (p<0.001). The authors also report that the baseline model produces error-free accompaniments more often (57.14%) than the FP model (45.71%), a tradeoff they attribute to the small fine-tuning set.

Load-bearing premise

The entire enterprise rests on the accuracy of one transcriber's reading of sometimes illegible or ambiguous manuscripts; if those readings are systematically wrong, the dataset misrepresents Price and the model learns a distorted style.

Editorial extensions

If this is right

  • Researchers can now analyze Price's harmony, chromaticism, whole-tone passages, key and meter changes, and Juba rhythms from computable symbolic files rather than from manuscripts.
  • The generative model offers a way to complete or sketch piano parts for Price's five incomplete songs and for unfinished works, with style transfer from a small corpus.
  • The baseline comparison shows that style capture and note-level correctness can diverge: the Price model is preferred stylistically while producing more perceived errors, so downstream users should treat outputs as raw material for a composer's editing.
  • The per-song metadata and section onsets make the dataset a testbed for style-conditioned generative music research and for studies of how much data a single-composer model needs.

Reading between the lines

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

  • Beyond the paper, the leave-one-out design implies the style preference is not explained by memorizing the tested song, but the two evaluators are also the people who transcribed the dataset and selected checkpoints; an external expert panel would clarify whether the preference generalizes.
  • The same recipe—take a general multi-track infilling model, fine-tune on a small public-domain composer corpus, compare against a genre baseline—could be applied to other composers whose manuscripts have recently entered the public domain, turning archival transcription into a scalable route to style-conditioned generation.
  • Because the dataset records the transcriber's judgment on illegible accidentals, a confidence or uncertainty annotation layer could turn transcription ambiguities into a research object rather than a silent correction.
  • A two-stage pipeline that uses the FP model for stylistic seeds and a human (or error-detection model) to repair hard errors would likely outperform either model alone on co-creative usability.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces the Florence Price Art Song Dataset, a manually transcribed digital collection of 112 of Price's songs for voice and piano, released in MuseScore, MusicXML, MIDI, and PDF formats. The authors also fine-tune a T5-based symbolic music model (the Florence Price Piano Accompaniment Generator) on this dataset to generate piano accompaniments for input melodies, and they evaluate the model in a blind listening experiment in which two evaluators compare accompaniments from the fine-tuned model against a baseline fine-tuned on the OpenScore Lieder Corpus. The reported results are that the FP model is preferred more often than the baseline on both Price melodies (82/137, p=0.026) and popular melodies (99/137, p<0.001). The paper also compares generated accompaniments for Price's 'Rhapsody' as a case study and reports an error-analysis table.

Significance. The dataset is a substantial and valuable contribution: it is the first publicly available symbolic-format digital collection of Price's vocal music, fills a genuine gap in music information retrieval and musicology, and the authors are careful to document the transcription pipeline, the copyright-driven exclusion of 17 songs, and the release of multiple file formats and audio-generation scripts. The leave-one-out training strategy for evaluating the FP model on Price melodies is a thoughtful design choice that avoids trivial memorization. The generative model itself is a useful case study in fine-tuning on a very small dataset. However, the headline empirical claim about perceived style similarity is not yet supported by the evidence as presented because the statistical analysis ignores the clustered structure of the data and the evaluators are the authors themselves; these are load-bearing issues for the paper's third contribution.

major comments (3)
  1. [§5, Table 1] The p-values in Table 1 are computed from a binomial test that treats each of the 137 responses as an independent Bernoulli trial. This assumption is not justified: the responses are clustered by 10 melodies (for each melody type), 7 generated accompaniments per model per melody, and 2 evaluators. Evaluations of different accompaniments to the same melody by the same listener are very likely correlated, so the effective sample size is far smaller than 137. The Price-melody result (82/137, p=0.026) is fragile and may not remain significant under a mixed-effects logistic regression or a cluster-robust test. The paper should report per-melody and per-evaluator win counts and re-analyze the data with an appropriate model; as it stands, the headline claim that the FP model is 'perceived as being reflective of Florence Price's style more frequently than a baseline' is not soundly established.
  2. [§5, Evaluation] The two evaluators are the first and third authors, who curated the dataset and know the study's hypothesis, and the second author generated the stimuli after performing checkpoint selection. Although the presentation was blind, the evaluators' prior exposure to the training corpus and their investment in the outcome create strong demand characteristics and expectancy effects that are not eliminated by blinding. The paper does not acknowledge this as a fundamental limitation in the abstract or conclusion, where the result is presented as definitive ('shows that accompaniments generated by our model are perceived...'). The authors should either recruit external listeners (for instance, musicians familiar with Price's style but not involved in the project) or explicitly reframe the evaluation as a preliminary self-assessment with all its caveats. This is load-bearing because contribution (3) rests entirely on this perceptual claim.
  3. [§3.2, Transcription fidelity] The dataset's accuracy rests on the first author's judgment in places where the manuscripts are illegible, and the two subsequent reviews were performed by the same research team. The paper states this limitation, but the contribution claim that this is 'the most complete digital collection of Florence Price's vocal music works to date' should be matched with a more prominent caveat that the transcriptions have not been independently verified and may contain systematic errors, especially in accidentals and rhythm. While the twice-reviewed process is a strength, it checks internal consistency rather than correctness. The authors could strengthen this by providing a small inter-transcriber agreement measure or by soliciting community corrections in a visible way. Even though the authors already acknowledge this, the abstract and introduction currently present the dataset without this caveat, which risks misleading readers who use the released files as ground truth.
minor comments (4)
  1. [§5, Table 2] Table 2 is difficult to read because the column headers are misaligned in the text ("Our Baseline Model Model") and the percentage row is run together ("45.71% 57.14%with no errors"). The numbers appear to be internally consistent (FP model 45.71% error-free; baseline 57.14%), but the formatting should be cleaned up.
  2. [§5, Missing responses] The paper states that 'Six out of 280 responses were missing' but does not specify how missing responses were handled. Since each row of Table 1 has n=137 rather than 140, the reader needs to know whether the missing responses were dropped per melody type and whether the two evaluators contributed equally to the missing data.
  3. [§5, Rhapsody case study] In the analysis of the Rhapsody accompaniments, the authors describe the FP model's B-flat major harmony as 'more surprising and expressive' than Price's own F-sharp major harmony and as 'encouraging.' This is a value judgment presented without supporting evidence; the baseline model's choice actually matches Price's original harmony. The case study would be more informative if it compared the FP model's output to Price's actual accompaniment as well as to the baseline.
  4. [§4, Fine-tuning epochs] The paper reports that the authors selected 130 epochs for the baseline and 100 epochs for the FP model by listening to outputs, but it does not provide the validation/training curves or a reproducible stopping rule. Given the paper's own call for research into stopping criteria, it would be helpful to make the listening-based selection procedure more explicit, for example by documenting how many candidate checkpoints were examined and how outputs were sampled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the listening comparison is an empirical, blind preference experiment, and the paper's self-citations are not load-bearing for the central claim.

full rationale

The paper's central claims are the release of a digital dataset and a blind listening comparison between accompaniments generated by the Florence Price fine-tuned model and a baseline model. The listening outcome is not derived from the training data or from any fitted parameter in a way that would make it equivalent to the model's inputs; it is an empirical preference count (82/137 for Price melodies, 99/137 for popular melodies) with binomial p-values, and the test stimuli were generated using a leave-one-out procedure. No fitted parameter is renamed as a prediction: the FP model is fine-tuned on transcriptions, not on the evaluation responses, and the evaluation responses do not feed back into the model. The self-citations present are [12], the co-author's in-progress Florence B. Price Works Catalog, used only for the contextual count of 134 documented songs, and [16,17], the second author's Composer's Assistant, used as the starting checkpoint for fine-tuning. Neither makes the headline claim true by construction, and neither functions as a 'uniqueness theorem' or an ansatz smuggled in via citation. The fact that the evaluators are the first and third authors, and that the second author selected checkpoints by listening, raises experimental-validity concerns (expectancy effects, correlated responses within melodies and raters), but those are not circularity under the definition used here: the preference outcome is not logically forced by the dataset or the fine-tuning procedure. The Table 2 text/table discrepancy (45% versus 57.14%) is a reporting error, not a circular step. No equation or definition reduces to itself, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are postulated; the FP model is a fine-tuned artifact of an existing architecture. The load-bearing inputs are the transcription fidelity, the base model's tokenization, and the evaluators' style judgments. The listed free parameters are training and sampling choices selected by listening rather than by an objective criterion.

free parameters (5)
  • Number of fine-tuning epochs for the FP model = 100
    Selected by listening to generated accompaniments (§4); the same kind of listening judgment is later measured in the official evaluation, creating a mild selection-on-outcome concern.
  • Number of fine-tuning epochs for the baseline model = 130
    Selected by listening (§4); the baseline is the comparison arm of the listening experiment.
  • Number of fine-tuning epochs for leave-one-out models = 100
    Each leave-one-out model used for Price-melody evaluation was trained for 100 epochs (§5), matching the main FP checkpoint.
  • Inference sampling parameters = temperature=1.0, rhythmic temperature=1.5, nucleus p=0.95
    Chosen to maintain output diversity during generation (§4); no sensitivity analysis is reported.
  • Per-song mood and style metadata labels = categorical values by song
    The happy/sad label, mood adjective, and style tag are the first author's subjective judgments (§3.2); they are released metadata, not model parameters, and do not feed the central claims.
assumptions (5)
  • domain assumption Manual transcriptions faithfully represent Price's manuscripts
    The dataset and the model both inherit the first author's judgments about illegible passages and 'perceived errors in the manuscripts' (§3.2).
  • domain assumption Composer's Assistant v2.1 is a valid base model for symbolic infilling
    The FP model is the second author's own prior release [16, 17], fine-tuned on the new corpus (§4); its tokenization and masking scheme define the space of possible accompaniments.
  • domain assumption The two author-evaluators' choices measure 'reflective of Price's style'
    The listening experiment uses the first and third authors as evaluators because external listeners familiar with Price's style were not recruited (§5).
  • standard math Binomially distributed independent responses
    Table 1's p-values assume independent trials per row; responses are clustered within 10 melodies and 2 evaluators, so the independence assumption is likely violated in the direction of overstating significance.
  • domain assumption The documented count of 134 songs is complete
    The count relies on the in-progress Florence B. Price Works Catalog [12] by co-author Shadle plus two library collections (§3.1); if the count is inaccurate, the coverage claim (112 of 134) shifts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Florence Price Art Song Dataset and Piano Accompaniment Generator." pith.science (2026). https://pith.science/paper/3EFZC3RB

@misc{pith2026250623130,
  author       = {Pith},
  title        = {Pith review of: The Florence Price Art Song Dataset and Piano Accompaniment Generator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3EFZC3RB}},
  note         = {Machine review of arXiv:2506.23130}
}
read the original abstract

Florence B. Price was a composer in the early 20th century whose music reflects her upbringing in the American South, her African heritage, and her Western classical training. She is noted as the first African-American woman to have a symphony performed by a major orchestra. Her music has recently received renewed attention from both the public and the research community, decades after her death. In addition to other genres, Price was a prolific composer for solo voice and piano. Music historians have documented the existence of 134 art songs and piano/voice arrangements for spirituals and folk songs written by Price. We release a digital catalog of 112 of these works in MuseScore, MusicXML, MIDI, and PDF format. We also use this dataset to fine-tune a symbolic music generation model to generate accompaniments to melodies, and we conduct a blind listening experiment that shows that accompaniments generated by our model are perceived as being reflective of Florence Price's style more frequently than accompaniments generated by a baseline model. We release our model as the Florence Price Piano Accompaniment Generator alongside our dataset.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages

  1. [1]

    The Florence Price Art Song Dataset and Piano Accompaniment Generator

    INTRODUCTION Florence B. Price (April 9, 1888–June 3, 1953) was a prolific composer of classical music during the first half of the twentieth century. Born and raised in Little Rock, Arkansas, she studied piano, organ, and composition at the New England Conservatory, where she graduated with honors in 1906. For the next two decades, she primarily lived in...

    work page Pith review arXiv 1953

  2. [2]

    RELA TED WORK Sheet music has been digitized into computer-readable formats in numerous ways, including ABC notation, music21 [6], MIDI, and MusicXML. Digitized classical music datasets include the OpenScore Lieder Corpus [7], which includes 1356 works of vocal music (art songs) by classical music composers (primarily in the Romantic Era), and the OpenSco...

    work page 2023

  3. [3]

    I Remember

    THE DA TASET 3.1 Basic information 134 different Florence Price art songs are documented to exist between the Florence B. Price Works Catalog [12], the Marian Anderson Collection at the Penn Libraries, and the Mullins Library of the University of Arkansas. The Mullins Library at University of Arkansas holds the largest collection of Price’s music manuscri...

    work page 2024

  4. [4]

    Given the small size of our dataset, it would be infeasible to train such a model from scratch

    THE FLORENCE PRICE PIANO ACCOMPANIMENT GENERA TOR In this section, we describe how we created a model, hereafter referred to as the FP model , for writing piano accompaniments that reflect the style of Florence Price. Given the small size of our dataset, it would be infeasible to train such a model from scratch. Therefore, we started with the Composer’s A...

    work page

  5. [5]

    hard” and “soft

    EV ALUA TION Due to the novelty of the dataset presented in this paper, it would be difficult to find external evaluators who are familiar with the style of the music in the dataset. Therefore, the first and third authors, who have listened to all of the music in the dataset, participated as evaluators in a blind listening experiment to assess the extent ...

    work page

  6. [6]

    Using the catalog, we have fine-tuned and released a symbolic music model to generate piano accompaniments reflective of Price’s compositional style

    CONCLUSION We have created and released a digital catalog of Florence Price’s works for solo voice and piano in MuseScore, MusicXML, MIDI, and PDF format. Using the catalog, we have fine-tuned and released a symbolic music model to generate piano accompaniments reflective of Price’s compositional style. We have conducted a listening experiment and conclud...

    work page

  7. [7]

    ETHICS STA TEMENT All of the music we have released in the Florence Price Art Song Dataset and used to train the Florence Price Piano Accompaniment Generator entered the public domain in the United States of America on Jan. 1, 2024. We obtained access to Price’s manuscripts with permission from the University of Arkansas David W. Mullins Library and the U...

    work page 2024

  8. [8]

    The “OpenScore String Quartet

    M. R. H. Gotham, M. Redbond, B. Bower, and P. Jonas, “The “OpenScore String Quartet” Corpus,” in Proceedings of the 10th International Conference on Digital Libraries for Musicology . Milan Italy: ACM, Nov. 2023, pp. 49–57. [Online]. Available: https://dl.acm.org/doi/10.1145/3625135.3625155

    arXiv 2023

Show all 45 references
  1. [9]

    C. Z. Huang and T. Cooijmans. (2012) JSB- Chorales-Dataset. [Online]. Available: https: //github.com/czhuang/JSB-Chorales-dataset

    2012

  2. [10]

    R. L. Brown, The Heart of a Woman: The Life and Music of Florence B. Price . Urbana, IL, USA: University of Illinois Press, 2020

    2020

  3. [11]

    Florence Price and the Politics of Her Existence,

    S. Ege, “Florence Price and the Politics of Her Existence,” American Music, vol. 39, no. 1, pp. 1–23, 2021

    2021

  4. [12]

    After Lost Scores Are Found in Abandoned House, Musicians Give Life to Florence Price’s Music,

    K. T. Steward, “After Lost Scores Are Found in Abandoned House, Musicians Give Life to Florence Price’s Music,” 2018, kUAR, University of Arkansas at Little Rock Public Radio. Accessed March 26, 2025. [Online]. Available: https: //www.ualrpublicradio.org/local-regional-news/20...

    2018

  5. [13]

    Music appreciation,

    A. Harrison, “Music appreciation,” 2021. [Online]. Available: https://library .achievingthedream.org/ alamomusicappreciation/chapter/art-song/

    2021

  6. [14]

    Schubert’s ’Unfinished’ Symphony completed by artificial intelligence,

    E. Davis, “Schubert’s ’Unfinished’ Symphony completed by artificial intelligence,” Classic FM , 2019. [Online]. Available: https://www.classicfm.com/composers/schubert/ unfinished-symphony-completed-by-ai/

    2019

  7. [15]

    music21: A Toolkit For Computer- Aided Musicology

    M. Cuthbert, C. Ariza, B. Hogue, and J. W. Oberholtzer, “music21: A Toolkit For Computer- Aided Musicology.” [Online]. Available: http:// web.mit.edu/music21/

  8. [16]

    The OpenScore Lieder Corpus,

    M. R. H. Gotham and P. Jonas, “The OpenScore Lieder Corpus,” in Music Encoding Conference Proceedings 2021, S. Münnich and D. Rizo, Eds. Humanities Commons, 2022, pp. 131–136

    2021

  9. [17]

    Composer’s Assistant 2: Interactive Multi-Track MIDI Infilling with Fine-Grained User Control,

    M. Malandro, “Composer’s Assistant 2: Interactive Multi-Track MIDI Infilling with Fine-Grained User Control,” in Proc. 25th Int. Society for Music Information Retrieval Conf., San Francisco, CA, USA, 2024, pp. 438–445

    2024

  10. [18]

    Accomontage: Accompaniment Arrangement Via Phrase Selection and Style Transfer,

    J. Zhao and G. Xia, “Accomontage: Accompaniment Arrangement Via Phrase Selection and Style Transfer,” in Proc. 22nd Int. Society for Music Information Retrieval Conf., Online, 2021, pp. 833–840

    2021

  11. [19]

    DeepBach: a Steerable Model for Bach Chorales Generation,

    G. Hadjeres, F. Pachet, and F. Nielsen, “DeepBach: a Steerable Model for Bach Chorales Generation,” in Proc. 34th Int. Conf. Machine Learning , ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 1362–1371

    2017

  12. [20]

    The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling,

    Y . Wu, J. Gardner, E. Manilow, I. Simon, C. Hawthorne, and J. Engel, “The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling,” arXiv preprint arXiv:2209.14458, 2022

    2022 arXiv

  13. [21]

    D. Shadle. (2023) Florence B. Price Works Catalog. [Online]. Available: https://fbpricecatalog.net/

    2023

  14. [22]

    F. B. Price, 44 Art Songs and Spiritual Arrangements , R. Heard, Ed. Fayetteville, AR, USA: ClarNan Editions, 2015

    2015

  15. [23]

    Price’s manuscript of I Remember, p

    and ProgGP [24], it is shown that transformer models Figure 1. Price’s manuscript of I Remember, p. 1. [25] can be fine-tuned on small datasets to train the models to write in specific musical styles. However, fine-tuning for style-conditioned accompaniment generation is a mor...

  16. [24]

    Beyond the years,

    K. Slack, “Beyond the years,” 2024, album. [Online]. Available: https://www .avie-records.com/ releases/beyond-the-years-karen-slack/

    2024

  17. [25]

    Music Transformer,

    C.-Z. A. Huang, A. Vaswani, J. Uszkoreit, I. Simon, C. Hawthorne, N. Shazeer, A. M. Dai, M. D. Hoffman, M. Dinculescu, and D. Eck, “Music Transformer,” in Int. Conf. Learning Representations, 2019

    2019

  18. [26]

    Composer’s Assistant: An Interactive Transformer for Multi-Track MIDI Infilling,

    M. Malandro, “Composer’s Assistant: An Interactive Transformer for Multi-Track MIDI Infilling,” in Proc. 24th Int. Society for Music Information Retrieval Conf., Milan, Italy, 2023, pp. 327–334

    2023

  19. [27]

    Florence Price: An Analysis of Select Art Songs with Text by Female Poets,

    C. Jobson, “Florence Price: An Analysis of Select Art Songs with Text by Female Poets,” Doctoral Dissertation, University of Miami, 2019. [Online]. Available: https://scholarship .miami.edu/ esploro/outputs/doctoral/Florence-Price-An-Analysis- of-Select/991031447224302976

    2019

  20. [28]

    MMM : Exploring Conditional Multi-Track Music Generation with the Transformer,

    J. Ens and P. Pasquier, “MMM : Exploring Conditional Multi-Track Music Generation with the Transformer,” arXiv preprint arXiv: 2008.06048, 2020

    2008 arXiv

  21. [29]

    Flexible Generation with the Multi-Track Music Machine,

    J. Ens and P. Pasquier, “Flexible Generation with the Multi-Track Music Machine,” in Extended Abstracts for the Late-Breaking Demo Session of the 21st Int. Society for Music Information Retrieval Conf. , Montréal, Canada, 2020

    2020

  22. [30]

    MIDI-GPT: A Controllable Generative Model for Computer-Assisted Multitrack Music Composition,

    P. Pasquier, J. Ens, N. Fradet, P. Triana, D. Rizzotti, J.- B. Rolland, and M. Safi, “MIDI-GPT: A Controllable Generative Model for Computer-Assisted Multitrack Music Composition,” 2025. [Online]. Available: https: //arxiv.org/abs/2501.17011

    2025 arXiv

  23. [31]

    MusIAC: An Extensible Generative Framework for Music Infilling Applications with Multi-level Control,

    R. Guo, I. Simpson, C. Kiefer, T. Magnusson, and D. Herremans, “MusIAC: An Extensible Generative Framework for Music Infilling Applications with Multi-level Control,” inArtificial Intelligence in Music, Sound, Art and Design. EvoMUSART 2022. Lecture Notes in Computer Science, ...

    2022

  24. [32]

    ShredGP: Guitarist style- conditioned tablature generation with transformers,

    P. Sarmento, A. Kumar, D. Xie, C. Carr, Z. Zukowski, and M. Barthet, “ShredGP: Guitarist style- conditioned tablature generation with transformers,” in Proceedings of the 16th International Symposium on Computer Music Multidisciplinary Research . Zenodo, Nov. 2023, p. 112–121....

    2023 doi

  25. [33]

    ProgGP: From guitarpro tablature neural generation to progressive metal production,

    J. Loth, P. Sarmento, C. Carr, Z. Zukowski, and M. Barthet, “ProgGP: From guitarpro tablature neural generation to progressive metal production,” in Proceedings of the 16th International Symposium on Computer Music Multidisciplinary Research. Zenodo, Nov. 2023, p. 122–133. [On...

    2023 doi

  26. [34]

    I Remember,

    F. Price, “I Remember,” Archival object MC 988a 2 (object ID 501224), Special Collections Department, University of Arkansas Libraries, Fayetteville, AR, box 12, Folder 27, Repository ID: 501224; accessed June 23, 2025. [Online]. Available: https://uark .as.atlas- sys.com/repo...

    2025

  27. [35]

    Florence B. Price’s Compositional Style in Song Settings of Paul Laurence Dunbar,

    D. J. H. Jr, “Florence B. Price’s Compositional Style in Song Settings of Paul Laurence Dunbar,” Master’s Thesis, University of Nevada, Las Vegas, 2025. [Online]. Available: https://digitalscholarship .unlv.edu/ cgi/viewcontent.cgi?article=4621&context= thesesdissertations

    2025

  28. [36]

    Catastrophic interference in connectionist networks: The sequential learning problem,

    M. McCloskey and N. J. Cohen, “Catastrophic interference in connectionist networks: The sequential learning problem,” ser. Psychology of Learning and Motivation, G. H. Bower, Ed. Academic Press, 1989, vol. 24, pp. 109–165. [Online]. Available: https://www .sciencedirect.com/sc...

    1989

  29. [37]

    Exploring the Limits of Transfer Learning with a Unified Text- to-Text Transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the Limits of Transfer Learning with a Unified Text- to-Text Transformer,” Journal of Machine Learning Research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  30. [38]

    Transformers: State- of-the-Art Natural Language Processing,

    T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. Rush, “Transformers: State- of-the-Art Natur...

    2020

  31. [39]

    The Jazz Transformer on the Front Line: Exploring the Shortcomings of AI- Composed Music Through Quantitative Measures,

    S.-L. Wu and Y .-H. Yang, “The Jazz Transformer on the Front Line: Exploring the Shortcomings of AI- Composed Music Through Quantitative Measures,” in Proc. 21st Int. Society for Music Information Retrieval Conf., Montréal, Canada, 2020, pp. 142–149

    2020

  32. [40]

    On Information and Sufficiency,

    S. Kullback and R. A. Leibler, “On Information and Sufficiency,” The Annals of Mathematical Statistics , vol. 22, no. 1, pp. 79 – 86, 1951. [Online]. Available: https://doi.org/10.1214/aoms/1177729694

    1951

  33. [41]

    Tradformer: A Transformer Model of Traditional Music Transcriptions,

    L. Casini and B. L. T. Sturm, “Tradformer: A Transformer Model of Traditional Music Transcriptions,” in Proceedings of the Thirty- First International Joint Conference on Artificial Intelligence, IJCAI-22, L. D. Raedt, Ed. International Joint Conferences on Artificial Intellig...

    2022

  34. [42]

    From Jigs and Reels to Schottisar och Polskor: Generating Scandinavian-like Folk Music with Deep Recurrent Networks,

    S. Mossmyr, E. Hallström, B. L. Sturm, V . H. Vegeborn, and J. Wedin, “From Jigs and Reels to Schottisar och Polskor: Generating Scandinavian-like Folk Music with Deep Recurrent Networks,” in 16th Sound and Music Computing Conference (SMC2019) , 2019

    2019

  35. [43]

    Discrete Diffusion Probabilistic Models for Symbolic Music Generation,

    M. Plasser, S. Peter, and G. Widmer, “Discrete Diffusion Probabilistic Models for Symbolic Music Generation,” in Proceedings of the Thirty- Second International Joint Conference on Artificial Intelligence, IJCAI-23, 8 2023, pp. 5842–5850

    2023

  36. [44]

    The Curious Case of Neural Text Degeneration,

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y . Choi, “The Curious Case of Neural Text Degeneration,” in Int. Conf. Learning Representations, 2020

    2020

  37. [2015]

    However, until now, most of the sheet music to these songs has remained unpublished

    More of her recently discovered songs have been recorded and released, including Karen Slack’s Grammy®- winning album Beyond the Years: Unpublished Songs of Florence Price [14], containing 19 unpublished songs by Price and released in 2024. However, until now, most of the shee...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.