Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling

In the realm of music AI, arranging rich and structured multi-track accompaniments from a simple lead sheet presents significant challenges. Such challenges include maintaining track cohesion, ensuring long-term coherence, and optimizing computational efficiency. In this paper, we introduce a novel system that leverages prior modelling over disentangled style factors to address these challenges. Our method presents a two-stage process: initially, a piano arrangement is derived from the lead sheet by retrieving piano texture styles; subsequently, a multi-track orchestration is generated by infusing orchestral function styles into the piano arrangement. Our key design is the use of vector quantization and a unique multi-stream Transformer to model the long-term flow of the orchestration style, which enables flexible, controllable, and structured music generation. Experiments show that by factorizing the arrangement task into interpretable sub-stages, our approach enhances generative capacity while improving efficiency. Additionally, our system supports a variety of music genres and provides style control at different composition hierarchies. We further show that our system achieves superior coherence, structure, and overall arrangement quality compared to existing baselines.

Paper

Similar papers

Peer review

Reviewer AQ7F7/10 · confidence 4/52024-07-06

Summary

This paper introduces a novel music AI system that creates multi-track accompaniments from a lead sheet by leveraging disentangled style factors to enhance context coherence, creativity, and computational efficiency. The proposed two-stage process begins with generating a piano arrangement using piano texture styles, followed by orchestrating multiple tracks by incorporating orchestral function styles. The system employs vector quantization and a multi-stream Transformer for modeling orchestration styles, significantly improving arrangement quality and providing flexible control over various music genres and compositional levels.

Strengths

1. The paper addresses the significant task of Multi-Track Accompaniment Arrangement, making notable improvements in the field. 2. The authors have diligently constructed a comprehensive framework in both the experimental design and methodology, contributing valuable insights and data to the field.

Weaknesses

1. The introduction lacks clarity, with key terms and relationships not sufficiently defined before introducing mathematical notations and formulations. This obscures the motivation and makes it difficult to understand, especially as the abstract outlines challenges related to coherency, creativity, and efficiency that are not clearly addressed in the introduction. 2. The contributions of the paper, as discussed between lines 46-60 and 85-94, appear to be incremental rather than significant, primarily building upon work from AccoMontage. 3. There is a consistent omission of motivation when new concepts are introduced. For instance, the motivation behind formulating the multi-track accompaniment arrangement task is not made clear in the introduction after discussing style representations. Similarly, the reasons behind specific mechanisms in sections 3 and 4 are not elaborated upon.

Questions

1. Why does concatenating style factors potentially contradict the structure of existing conditions as mentioned on line 25? 2. The paper discusses conflicts when combining style factors in sequence generation tasks; could these be resolved using different methods like diffusion models? 3. How do the initial discussions on introducing style into long-term generation relate to the paper’s main theme of multi-track accompaniment arrangement? 4. What prompted the formulation of the problem between lines 32-36 in paragraph 3, and how is time t determined? 5. Is there a distinction between 'piano reduction' and 'piano track'? The terminology varies between italics and normal text; does this signify a difference? 6. What is the rationale behind using VQ-VAE and VAE to encode orchestral functions and piano reductions in section 3.2? 7. What motivations underlie the construction of the Multi-Track Orchestral Function Prior in section 3.3? 8. Can you explain what inputs the proposed model receives and provide comparisons to highlight its unique features, as seen in Figure 4? 9. How does the model handle both ABC notation and Midi, and is there a conversion of ABC notation into the representation discussed in section 3.1? 10.Why is the computational efficiency of this model less optimal compared to [20]? 11. How are the 'style' factors controlled and tested in the experiments?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are addressed.

Reviewer ANaR6/10 · confidence 4/52024-07-10

Summary

The paper presents a style transfer-based music accompaniment generation system. It starts with the lead melody and fleshes out the accompaniment tracks based on various high-level information, such as instrument type and rhythmic structure. The main goal of the proposed model is to be able to generate coherent structure in long music sequences, while they are also cohesive. Model architecture adopts VQVAE baselines to learn latent representations while the sequence learning part is done via transformer models applied to tracks and time dimensions orthogonally. The objective and subjective test scores look promising.

Strengths

- The paper presents reasonable intermediate representations of symbolic music, piano reduction and orchestral function, which are used to learn discrete embeddings by the VQVAE model. - The proposed model is evaluated by both objective and subjective manners, where it outperformed the other methods.

Weaknesses

- The description of the proposed model is not organized well, so it is difficult to understand. - Some high-level choices the authors made are not well justified, such as discrete embeddings vs. continuous, Gaussian noise vs. other noise, etc.

Questions

- In Fig. 1, the autoencoder takes the piano reduction information as its first input. I imagine it would correspond to the lead sheet. Does it mean that during the training time the output of the autoencoder is the full sheet with all tracks? Then this isn't an autoencoder, right? The precise training procedures are not detailed enough in this manuscript. - If the input PN is the full music during training, how does the model generalize to the test-time scenario when the input has to be only the lead sheet? - The manuscript seems to be based on the concept of 1 segment = 8 beats. Is this correct? Any rationale behind this choice? - Gaussian noise is used to regularize the model. But if it's to deal with the domain shift between piano reduction representations, wouldn't there be a different type of noise that's more suitable, i.e. more discrete random variables? - It appears that the "s" embeddings contain important information about the track to be generated. Since it's learned in an unsupervised way, it's a little hard to imagine how these embeddings are more interpretable than other existing conditioning vectors as the authors claim.

Rating

6

Confidence

4

Soundness

4

Presentation

2

Contribution

3

Limitations

- Basically it's not clear how Fig. 1 and Fig. 2 are combined to form the pipeline described in Fig 3. Colored figures, shades, and hashes show the authors' effort in describing this method, but it's not entirely clear to me. - It appears that the system takes the user input to designate which instrumental track to generate based on the instrument embedding. It's a nice feature, but it also means that, for evaluation, somebody has to come up with a nice orchestration (i.e., which instruments go well given the lead melody). Was this critical selection of instruments done by the authors to generate the test sequences? Unless I missed this part, it has to be clearly mentioned, as it can affect the performance of the model (in comparison to other methods that don't have such a concept).

Reviewer Y3LG7/10 · confidence 4/52024-07-13

Summary

The paper suggests creating multi-instrument accompaniment by using the piano reduction and the instrument note density (referred to as the 'Orchestration Function') as bootstrap representations. By effectively applying VAE and an autoregressive sequence generation framework, the paper demonstrates the high potential and effectiveness of the proposed approach for learning musically structured accompaniment. The paper also proposed a 'layer interleaving architecture' that processes the orchestration codec by alternating between the time axis and the track axis. The paper compares with previous works with objective and subjective evaluation and shows its validity.

Strengths

The strength of this paper, I believe, lies in the factorization of accompaniment generation by leveraging the fact that piano reduction and 'orchestration function' can be freely produced as middle-level features from multi-track MIDI. By separating each process, as the authors claim, it becomes applicable to scenarios where users can exert control. Additionally, by dividing the modeling of each accompaniment's prior and the modeling of detailed notes, the efficiency of learning for each model is expected to have increased.

Weaknesses

The proposed methodology utilizes piano reduction as a condition, so the quality of the final orchestra accompaniment is expected to vary depending on the quality of the piano reduction. Upon listening to the provided examples, it is evident that notes not present in the piano reduction were added, confirming that the proposed model's role extends beyond merely rearranging the notes of the piano reduction. However, there needs to be a discussion on the impact of using piano reduction on the model's high evaluation compared to other models. Additionally, while it was mentioned in L168 that another module was used, a more detailed explanation of how the piano reduction was generated is necessary.

Questions

Q1. From my understanding, the 'orchestration function' in Eq. 1 eliminates pitch information. Thus, I expect the priors (${s_t}$) to provide note-density-like bare information. What do you expect to be encoded in $s$? Also, as a follow-up question, what kinds of correlation are expected along with $s_t$, and why does it make sense to train a sequential prediction model for $s$? I'm not sure about the expected role of the prior model.

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

I generally agree with the authors about the limitation presented in Appendix E.

Reviewer iTts8/10 · confidence 4/52024-07-16

Summary

The authors introduce a new model for generating multi-track accompaniment given a lead sheet. While being strictly a two-stage method (first generate a piano arrangement from the lead sheet and *then* generate the accompaniment), the authors' contribution is focused on the 2nd stage, and they rely on existing modules for the 1st stage. Their proposed method features two key components: **first**, a VQ-VAE submodule whose aim is to learn (quantized) representations of "orchestral functions" (orchestral functions being 1-D time-series representations of a track, with each element representing the sum of notes active in each segment, more on that in weaknesses), and, **second**, a transformer-based "mixing" VAE which combines the representations of the VQ-VAE (i.e., the representations of the orchestral functions) with learnt representations of "piano reductions" of each song (derived by averaging all individual tracks) to generate latent representations which are used to synthesize the full song. A key novelty of their model is the "mixing" part of the VAE, with the decoder utilizing interleaved time-wise and track-wise cross-attention layers -- meaning, that each output "sees" information both from past frames in its own track and from other tracks. Their system compares favorably to 3 standard baselines taken as-is from prior work or reproduced.

Strengths

Overall, barring some presentation issues discussed below, this is a well-written paper that features state-of-the-art results and enough novelty to make it relevant for both the broader generative AI community (given the difficulty of generating long-form, complex music which the authors tackle with hierarchical modelling) and the more niche music-generation community. This is why I am leaning positive towards accepting the paper. I will mention some more of its strengths in detail: + The authors target a challenging task which is to generate a multi-track arrangement given a lead sheet. Even though they "only" focus on the 2nd stage of that pipeline, namely the generation of the multi-track arrangement given a piano arrangement, it is still a valid contribution. + The idea to disentangle the piano reduction and orchestral function and combine them in the way they do is interesting and novel. This also provides a nice element of controllability. + The fact that the model can be trained in "self-supervised" fashion (i.e. without manual annotations) is commendable. + The authors have done a thorough job of documenting experimental parameters (data, GPU runtime, etc.) and submit code to reproduce their experiments. + The authors have accounted for different sources of variability and provided multiple "knobs" to twist (and counteract domain mismatches) through the clever use of multiple "positional" embeddings (Fig. 2).

Weaknesses

The main weaknesses of the paper are primarily in terms of presentation, rather than methodology or evaluation. My comments are mainly meant to help the authors improve their presentation. * Clarity of music concepts: Sometimes, the authors introduce or reference concepts which are not straightforward without domain knowledge. Examples of that are: * The orchestral function of x is introduced as a column sum of activation indicators over the different columns of X. Given that the columns of X represent MIDI pitches, this would mean that at each frame, the authors are summing the notes that are active. Yet, in p.3, l.115, they write that this indicator function "counts each note onset position as 1". Why onset only? This would mean that each note is counted once. But the equation states $x_t^k>0$, which appears to be evaluated for each frame (thus, not only at the onset of each note, but for its entire duration). Further, they state that "it [orchestral function] describes the rhythm and groove of each track". It is not at all clear to me why this orchestral function should describe the rhythm? It only appears to describe the amount of notes active at a given time. It's even less clear why it should describe the groove, given the subjective nature of "groove". It is okay that the authors attempt to give a layman's explanation to their method in this part, but it is also recommended that they better describe those concepts (and do so in a way that would be acceptable even to those that are not familiar with concepts from music theory). * There are other terms, such as "counterpoint", "phrase", etc. which will not be evident in an audience not familiar with them. While it is understandable that in a music generation paper, there will be a lot of domain language, it would be useful if the authors commented on their importance in their discussion of results. For example, the ILS metric measures the similarity within a music phrase vs the rest of the song. What does it mean that the authors' method does better than the baselines in this metric? That it produces more coherent phrases that are more easily distinguishable from the rest of the song? Adding a sentence to clarify the importance of each result (in, as far as this is possible, layman's terms) would improve the readability and outreach of the paper * Comparability to baselines: The authors compare to three established baselines. However, with the exception of PopMAG, which the authors reproduce on part of the data they used to train their own model, the other two baselines are not strictly comparable. AMT requires ground-truth accompaniment (arguably a limitation of that method compared to the authors'), and GETMusic was trained on a different data (and additionally handles different instrumental tracks). While this is not a no-go, and is common practice for generative methods in general, it should be mentioned in 5.2. * The description of *interleaving* layers in the autoencoder -- according to the authors themselves, a major novelty of the work -- does not feature as prominently in their text as it should (given its stated importance), and could be improved in terms of clarity. Specifically, it is not entirely clear what the "Track Encoder" does in Fig. 2. Is it a standard residual "transformer" (it is not clear from A.2 if a residual connection is there) layer that essentially combines information across tokens? It would be beneficial to describe this in 3.3 using simple equations like in 3.2. One important open question is why this track encoder, if it is indeed based on self-attention, is different than the decoder? Specifically, why does one integrate information in the time-axis and the other on the track-axis? It would be important to show the dimensions here. I guess you have: $(B, K, N, P)$, where B is the batch size, K the number of tracks, N the duration of each note, and P the number of notes (I guess N&P are eventually downsampled from their original dimensions). So the inter-track encoder would operate on tensors $(B\cdot N, K, P)$ and then the intra-track decoder would operate on tensors $(B\cdot K, N, P)$? All this of course would be done in autoregressive fashion, so it would go from $\{1...N\}$. It would be nice to show this using the notation of the authors. * The VQ-VAE is described (A.1) and portrayed (Fig. 1) as having a convolutional encoder with a stride of 4 (thus a downsampling of 4) and a fully-connected decoder. It would be useful to show how this fully-connected decoder mitigates the downsampling of the encoder to have an output size equal to the input. Minor comments: + p.7, l.240: Best substitute "denoising process" with denoising diffusion probabilistic model (if this is what this meant), as the term denoising can be ambiguous

Questions

There is only one question, regarding the interleaving of layers, but since it is also a weakness, I have included it there instead.

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

The authors have adequately discussed their limitations.

Reviewer AQ7F2024-08-13

Thank you for your detailed and thoughtful response to my review. Your clarifications, particularly regarding the significance of your contribution and the motivations for the methodologies employed, provide a deeper understanding of your work. I look forward to seeing these enhanced explanations incorporated into the revised manuscript, as they will undoubtedly strengthen its clarity and justification.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC