Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Singing voice AI can't tell rock from pop without retraining

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 22:07 UTC pith:34MBBCE2

load-bearing objection First multi-genre SVS benchmark; finds genre collapse is data-dependent, not controllable at inference the 4 major comments →

arxiv 2607.06986 v1 pith:34MBBCE2 submitted 2026-07-08 cs.SD

MMGenre: Benchmarking Singing Voice Synthesis across Multiple Musical Genres

Wenhao Feng , Yuxun Tang , Jiatong Shi , Qin Jin This is my paper
classification cs.SD
keywords genresmmgenresynthesisacrosssingingvoiceanalysisevaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces MMGenre, a benchmark spanning 10 musical genres and 26 subgenres, designed to test whether singing voice synthesis (SVS) systems can produce genre-appropriate vocals. The authors build the benchmark using an automatic pipeline: a text-to-music generator creates genre-conditioned audio, vocal tracks are separated, and symbolic music scores are extracted. When eight representative SVS models are evaluated on this benchmark, a striking pattern emerges: regardless of the target genre, all models produce vocals that sound acoustically similar to pop music. The genre-discriminative information present in the input scores is effectively lost during synthesis. The authors call this phenomenon 'genre collapse.' They then investigate whether genre awareness can be recovered through inference-time control strategies (style transfer, technique conditioning) or through targeted training. Zero-shot inference-time strategies yield only marginal improvements. However, just two hours of genre-specific continued training on Rock raises the genre consistency score from 1.5 to 4.9, nearly matching ground truth. This demonstrates that genre awareness in current SVS systems is not an emergent capability of the model architecture or the symbolic score representation, but is entirely dependent on the distribution of the training data.

Core claim

Current singing voice synthesis systems exhibit genre collapse: they produce pop-like vocals regardless of the target genre, and this behavior cannot be fixed through inference-time conditioning but can be substantially corrected through even limited genre-specific fine-tuning, proving that genre awareness is a distribution-dependent property rather than an emergent model capability.

What carries the argument

The central mechanism is the interplay between symbolic score representation and training data distribution. The paper disentangles whether genre-discriminative information resides in the phoneme-pitch-duration score (which would enable inference-time control) or in the learned statistical priors of the training data (which would require retraining). The finding that zero-shot control fails while targeted training succeeds isolates training data composition as the dominant factor governing genre behavior in SVS.

Load-bearing premise

The benchmark's validity rests on the claim that AI-generated singing audio preserves the relative ranking of SVS systems the same way real singing does. This equivalence was verified only for Pop music, so it remains unconfirmed whether the benchmark's synthesized singing interacts with SVS models identically in non-Pop genres.

What would settle it

If a future SVS model architecture were designed with explicit genre conditioning that successfully produced genre-discriminative vocals through inference-time control alone (without genre-specific training data), the claim that genre awareness is purely distribution-dependent would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • SVS benchmarks should evaluate genre diversity as a first-class dimension, not just overall naturalness or intelligibility, to avoid masking systematic stylistic collapse.
  • The genre collapse phenomenon likely extends to other style attributes in generative audio systems, suggesting that controllability claims based on inference-time conditioning may overstate what models can do without distributional changes.
  • The automatic pipeline for constructing genre-aligned score-audio pairs from text-to-music systems offers a scalable method for building controlled evaluation data in domains where real-world multi-genre datasets are scarce or biased.
  • Improving genre awareness in SVS may require explicit genre conditioning mechanisms in model architectures rather than relying on implicit emergence from score representations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If genre collapse occurs in SVS due to training data bias toward pop, analogous collapse phenomena likely exist in text-to-speech systems trained on narrow demographic or stylistic distributions, producing a default speaker identity that overrides conditioning signals.
  • The finding that Suno-generated singing preserves relative SVS model rankings (validated only on Pop) raises the question of whether text-to-music systems themselves exhibit genre collapse, which could propagate into benchmarks built from their output.
  • The dramatic improvement from two hours of Rock-specific training suggests that genre-specific vocal features may occupy a compact, learnable subspace, raising the possibility that efficient multi-genre SVS could be achieved through lightweight adapter modules rather than full multi-genre datasets.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces MMGenre, a benchmark for evaluating singing voice synthesis (SVS) across 10 musical genres and 26 subgenres. The authors construct genre-aligned score–audio pairs using an automatic pipeline (Suno V4.5 for music generation, Mel-RoFormer for vocal separation, STARS for score annotation). The benchmark is used to evaluate 8 representative SVS models, revealing a

Significance. The paper addresses a genuine gap in SVS evaluation: existing benchmarks are overwhelmingly Pop-centric, and genre has not been treated as a first-class evaluation dimension. The diagnostic finding of genre collapse — that SVS models produce acoustically similar outputs across genres regardless of input genre — is well-supported by the multi-model, multi-genre evaluation (Fig. 4, Fig. 5a) and is valuable to the community. The pipeline design (using T2M generation to construct controlled genre-balanced data) is a practical and reusable contribution. The benchmark validity analysis (§2.4, Spearman ρ=0.90) provides reasonable evidence that synthesized singing preserves relative SVS model rankings, though the scope of that validation is limited (see major comments). The use of an external automatic rater (Gemini 2.5 Pro) with a human agreement check (ρ=0.85) is a reasonable methodological choice for scalable evaluation.

major comments (4)
  1. §3.3, Training-Time Distributional Bias: The fine-tuning experiment — the strongest evidence for the claim that genre awareness is distribution-dependent rather than score-conditioned — is conducted on a single genre (Rock). Rock is acoustically distinctive from Pop (the paper itself notes 'broadband high-frequency energy, sharp transient structures, and irregular harmonic spreads' in Fig. 5b), which may make it an unusually easy case for fine-tuning to produce perceptible genre shifts. The zero-shot experiment (Table 2) tests three genres (Classical, Rock, Rap), but the fine-tuning experiment tests only Rock, creating an asymmetry in the evidential chain. If fine-tuning on a genre with subtler vocal differences from Pop (e.g., Jazz phrasing, Classical art-song) yielded smaller gains, the claim that genre awareness is purely distribution-dependent would require qualification. The authors
  2. should either (a) extend the fine-tuning experiment to at least 2–3 additional genres spanning a range of acoustic distinctiveness from Pop, or (b) explicitly acknowledge this limitation and temper the conclusion in §4 ('genre behavior in current SVS systems is primarily governed by learned distributional priors') accordingly.
  3. §3.1.2, GCS-5 metric: The Genre Consistency Score is the central evaluation metric for all findings, yet the prompt design for Gemini 2.5 Pro is not specified in the paper (it is listed as a free parameter). The human agreement check (ρ=0.85) covers 100 samples across 5 genres, but the reliability across all 26 subgenres — particularly for genres where genre boundaries are perceptually ambiguous (e.g., R&B vs. Soul subgenres, Blues vs. Country) — is not established. Given that GCS-5 scores are used to support the genre collapse finding across all genres, the authors should provide the prompt template and discuss potential rater biases or sensitivities, especially for acoustically adjacent genres.
  4. §2.4, Benchmark Validity: The validation that Suno-synthesized singing preserves relative SVS model rankings (ρ=0.90) is restricted to Pop only. While the authors note that SVS models consume extracted symbolic scores rather than Suno audio directly (which mitigates the concern), the score extraction pipeline (STARS) could still propagate genre-specific artifacts from Suno's generation into the symbolic representation (e.g., systematic pitch range or duration patterns that differ from real-world genre scores). The score-level diversity analysis (Fig. 3b) shows pitch distribution differences across genres, but it is unclear whether these reflect genuine genre characteristics or Suno's generation biases. A brief discussion of this potential confound, or a validation on at least one non-Pop genre, would strengthen the benchmark's reliability.
minor comments (6)
  1. The benchmark is Chinese-only (3,152 Chinese score–audio pairs). This is not prominently acknowledged as a limitation. Genre conventions (especially for Rap, Country, Blues) may manifest differently across languages, and the generalizability of findings to other languages is unclear.
  2. §3.3: The fine-tuned model achieves GCS-5 = 4.9, slightly higher than ground truth (4.8). The authors attribute this to 'subgenre-specific exaggeration of salient Rock cues,' but this is unverified. A brief analysis or listening test would strengthen this interpretation.
  3. The fine-tuning data (2 hours of AI-generated Rock data) is described as 'independently constructed' and 'disjoint from the benchmark set,' but the construction methodology and genre consistency verification for this training data are not described.
  4. Fig. 4: The radar plot is difficult to read due to the density of overlapping lines. Consider using a heatmap or faceted bar chart for clearer comparison.
  5. Suno V4.5 is a proprietary system, which affects reproducibility of the pipeline. The authors mention modularity ('allows straightforward extension to new genres, languages, or alternative text-to-music generators') but do not discuss sensitivity to the choice of T2M model.
  6. Table 1: The CER values are quite high (0.25–0.69), suggesting significant lyric intelligibility issues across all models. This is worth brief discussion, as it may interact with genre perception.

Circularity Check

0 steps flagged

No significant circularity; central claim derived from independent external evaluation

full rationale

The paper's central claim — that genre awareness in SVS is distribution-dependent rather than an emergent property of score conditioning — is derived from evaluating external SVS models (RNN, XiaoiceSing, VISinger, DiffSinger, etc.) on the benchmark, using an external rater (Gemini 2.5 Pro) validated against human judgment (ρ=0.85). The benchmark construction pipeline uses external tools (Suno, Mel-RoFormer, MuQ-MuLan) alongside self-authored tools (STARS, SingMOS, SingMOS-Pro), but these self-citations are tool usage, not load-bearing justifications for the central claim. The fine-tuning experiment (§3.3) uses training data 'disjoint from the benchmark set,' preventing trivial circularity. The benchmark validity check (§2.4) validates synthesized singing against real singing with an independent MOS study (ρ=0.90). One minor near-circularity: MuQ-MuLan is used both for genre-consistency filtering during data construction (§2.2) and for the diversity analysis showing genre separability (§2.3, Fig 3a) — the high recognition accuracy (76% Top-1) could be partly inflated by the filtering step. However, this diversity analysis is not load-bearing for the central genre-collapse claim, which rests on GCS-5 scores and SVS output embeddings, not on the input data's separability. No derivation step reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The paper introduces two main invented entities (MMGenre benchmark and GCS-5 metric) and relies on three key axioms. The benchmark has independent evidence through its validity check and public availability. The GCS-5 metric has independent evidence through human correlation. The weakest axiom is that the Pop-only validity check generalizes to all genres, which is an ad-hoc assumption specific to this paper's validation scope.

free parameters (3)
  • GCS-5 prompt design
    The exact prompt given to Gemini 2.5 Pro for genre consistency scoring is not specified in the paper. The prompt design influences the absolute GCS-5 values.
  • Suno prompt design
    Textual prompts for Suno V4.5 are 'initially assisted by ChatGPT, followed by manual refinement.' The specific prompts are not listed, introducing variability in the generated benchmark data.
  • Fine-tuning data volume (2 hours) = 2 hours
    The Rock fine-tuning experiment uses 2 hours of AI-generated data. This quantity is chosen ad hoc; the paper does not explore sensitivity to this parameter.
axioms (3)
  • domain assumption Suno V4.5 generates genre-conditioned music that is sufficiently genre-consistent for benchmark construction.
    The benchmark relies on Suno producing genre-distinct music. This is partially validated by MuQ-MuLan recognition accuracy (Top-1: 76%, Top-3: 92%) in §2.3, but the axiom is that this accuracy is sufficient for the benchmark's purpose.
  • domain assumption Gemini 2.5 Pro's genre consistency ratings are a reliable proxy for human perception of genre alignment.
    The GCS-5 metric is the core evaluation metric. Human agreement check on 100 samples shows ρ=0.85, but the paper assumes this correlation holds across all genres and all model outputs.
  • ad hoc to paper The relative ranking of SVS models on synthesized Pop singing generalizes to non-Pop genres.
    The benchmark validity check (§2.4) is restricted to Pop. The paper assumes that the synthesized singing source does not differentially affect SVS evaluation in other genres.
invented entities (2)
  • MMGenre benchmark independent evidence
    purpose: A standardized framework for evaluating genre generalization in SVS across 10 major genres and 26 subgenres.
    The benchmark is constructed via a reproducible pipeline and its validity is checked against real singing data with Spearman ρ=0.90. The benchmark data and code are stated to be available via the project page.
  • GCS-5 (Genre Consistency Score) independent evidence
    purpose: A 5-point perceptual rating measuring how well synthesized singing matches the target genre, using Gemini 2.5 Pro as rater.
    The metric is validated against human ratings (ρ=0.85 on 100 samples). It provides a falsifiable, scalable proxy for genre alignment.

pith-pipeline@v1.1.0-glm · 12568 in / 2714 out tokens · 271205 ms · 2026-07-09T22:07:25.200943+00:00 · methodology

0 comments
read the original abstract

Singing voice synthesis (SVS) has progressed rapidly, yet its ability to generalize across diverse musical genres remains underexplored. Existing benchmarks are heavily biased toward pop music, limiting systematic analysis of genre-dependent behavior. We introduce MMGenre, a benchmark for multi-genre SVS diagnosis, supported by an automatic pipeline for constructing genre-aligned music scores. MMGenre spans 10 major genres and 26 subgenres, enabling comprehensive analysis of genre-aware synthesis. Extensive evaluation of representative SVS models reveals limited genre discrimination: synthesized vocals across genres exhibit highly similar acoustic characteristics and weak separability. While zero-shot genre adaptation yields only marginal improvements, lightweight genre-specific continued training leads to substantial gains. MMGenre provides a standardized framework for multi-genre SVS evaluation and exposes critical challenges in achieving genre-aware singing voice synthesis.

Figures

Figures reproduced from arXiv: 2607.06986 by Jiatong Shi, Qin Jin, Wenhao Feng, Yuxun Tang.

Figure 2
Figure 2. Figure 2: Benchmark data construction pipeline. The final benchmark data consist of genre-conditioned singing voice and aligned symbolic music score pairs. ments with abnormal durations are removed, and genre consis￾tency is automatically verified using MuQ-MuLan [28]. A small portion of samples is then checked by human listeners to elim￾inate occasional artifacts introduced by automatic processing. The resulting si… view at source ↗
Figure 3
Figure 3. Figure 3: Diversity analysis from audio and symbolic perspec￾tives. (a) UMAP visualization of MuQ audio embeddings for genre-conditioned Suno-generated samples. (b) Distribution of MIDI note pitches across selected genres. 3. Experiments The central question of this study is whether current SVS mod￾els can generate genre-consistent singing voices from genre￾specific musical scores? To systematically investigate this… view at source ↗
Figure 4
Figure 4. Figure 4: Genre-wise GCS-5 radar plot across SVS models. All models exhibit highly similar genre profiles, with strong align￾ment concentrated in Pop-related genres, revealing a systematic genre collapse phenomenon. 3.2. Benchmark Results 3.2.1. Genre-wise Alignment Results We report genre alignment metrics of different SVS systems across all genre categories. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the genre collapse phenomenon. (a) UMAP projection of MuQ embeddings for SVS inference audio, showing substantial cross-genre overlap. (b) Log-mel spectro￾gram comparison on Rock across three representative samples (columns 1–3) [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 10 internal anchors

  1. [1]

    MMGenre: Benchmarking Singing Voice Synthesis across Multiple Musical Genres

    Introduction Singing voice synthesis (SVS) aims to generate expressive singing voices directly from symbolic music scores and has achieved substantial progress with recent neural and diffusion- based models [1–4]. Prior research has largely focused on im- proving naturalness [5–7], expressiveness [8], and controllabil- ity [9, 10] of synthesized singing v...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Metal rock style, male vocal, distorted guitars, fast tempo

    Multi-Genre Benchmark 2.1. Benchmark Overview MMGenre is designed to provide a unified benchmark for evalu- ating SVS under diverse musical genre settings. The benchmark covers 10 major musical genres and 26 representative subgenres organized in a hierarchical taxonomy. As illustrated in Fig. 1, each major genre is associated with 2–3 representative subge...

    work page

  3. [3]

    Experiments The central question of this study iswhether current SVS mod- els can generate genre-consistent singing voices from genre- specific musical scores?To systematically investigate this is- sue, we conduct a structured empirical analysis that examines genre behavior from perceptual, representational, and training- dynamics perspectives. 3.1. Exper...

    work page

  4. [4]

    As shown in Fig. 5(b), the fine-tuned model exhibits stronger high-frequency energy and sharper transient structures, closer to ground-truth Rock vocals, while the baseline output retains smoother, Pop-like characteristics. The substantial improvement indicates that genre- consistent vocals can emerge with sufficient genre-specific training data. This con...

    work page

  5. [5]

    Conclusion We present MMGenre, a unified benchmark for evaluating genre-aware singing voice synthesis across multiple musical genres under standardized and controlled conditions. Through systematic evaluation, we reveal a consistent genre collapse phenomenon: current SVS systems exhibit limited genre dif- ferentiation and tend to converge toward a dominan...

    work page

  6. [6]

    62576347)

    Acknowledgments This work was partially supported by the National Natural Sci- ence Foundation of China (No. 62576347). We thank Renke Huang for insightful discussions and suggestions on music- related aspects of this work, particularly on the design of the genre hierarchy

    work page

  7. [7]

    All technical content, ex- perimental design, results, and conclusions were produced and verified by the authors

    Generative AI Use Disclosure Generative AI tools were used solely for language polishing and improving the clarity of expression. All technical content, ex- perimental design, results, and conclusions were produced and verified by the authors

    work page

  8. [8]

    Muskits: an End-to-end Music Processing Toolkit for Singing V oice Synthesis,

    J. Shiet al., “Muskits: an End-to-end Music Processing Toolkit for Singing V oice Synthesis,” inProc. Interspeech, 2022

    work page 2022

  9. [9]

    Sequence-to- sequence singing voice synthesis with perceptual entropy loss,

    J. Shi, S. Guo, N. Huo, Y . Zhang, and Q. Jin, “Sequence-to- sequence singing voice synthesis with perceptual entropy loss,” inICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 76–80

    work page 2021

  10. [10]

    Visinger: Variational inference with adversarial learning for end-to-end singing voice synthesis,

    Y . Zhang, J. Cong, H. Xue, L. Xie, P. Zhu, and M. Bi, “Visinger: Variational inference with adversarial learning for end-to-end singing voice synthesis,” inICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7237–7241

    work page 2022

  11. [11]

    Diffsinger: Singing voice synthesis via shallow diffusion mechanism,

    J. Liu, C. Li, Y . Ren, F. Chen, and Z. Zhao, “Diffsinger: Singing voice synthesis via shallow diffusion mechanism,” inProceedings of the AAAI conference on artificial intelligence, vol. 36, no. 10, 2022, pp. 11 020–11 028

    work page 2022

  12. [12]

    XiaoiceSing: A high-quality and integrated singing voice synthesis system,

    L. Peilinget al., “XiaoiceSing: A high-quality and integrated singing voice synthesis system,”Proc. Interspeech, 2020

    work page 2020

  13. [13]

    VISinger2: High-Fidelity End-to-End Singing V oice Synthesis Enhanced by Digital Signal Processing Synthe- sizer,

    Y . Zhanget al., “VISinger2: High-Fidelity End-to-End Singing V oice Synthesis Enhanced by Digital Signal Processing Synthe- sizer,” inProc. Interspeech, 2023

    work page 2023

  14. [14]

    TokSing: Singing voice synthesis based on discrete tokens,

    Y . Wuet al., “TokSing: Singing voice synthesis based on discrete tokens,” inProc. Interspeech, 2024

    work page 2024

  15. [15]

    Expressivesinger: Multilingual and multi-style score-based singing voice synthesis with expressive performance control,

    S. Daiet al., “Expressivesinger: Multilingual and multi-style score-based singing voice synthesis with expressive performance control,” inProc. ACM MM, 2024

    work page 2024

  16. [16]

    Stylesinger: Style transfer for out-of- domain singing voice synthesis,

    Y . Zhang, R. Huang, R. Li, J. He, Y . Xia, F. Chen, X. Duan, B. Huai, and Z. Zhao, “Stylesinger: Style transfer for out-of- domain singing voice synthesis,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 17, 2024, pp. 19 597–19 605

    work page 2024

  17. [17]

    Techsinger: Technique controllable mul- tilingual singing voice synthesis via flow matching,

    W. Guo, Y . Zhang, C. Pan, R. Huang, L. Tang, R. Li, Z. Hong, Y . Wang, and Z. Zhao, “Techsinger: Technique controllable mul- tilingual singing voice synthesis via flow matching,” inProceed- ings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 22, 2025, pp. 23 978–23 986

    work page 2025

  18. [18]

    The science of singing voice,

    J. Sundberg and T. D. Rossing, “The science of singing voice,” 1990

    work page 1990

  19. [19]

    The do re mi’s of everyday life: the structure and personality correlates of music preferences

    P. J. Rentfrow and S. D. Gosling, “The do re mi’s of everyday life: the structure and personality correlates of music preferences.” Journal of personality and social psychology, vol. 84, no. 6, p. 1236, 2003

    work page 2003

  20. [20]

    Utilizing human memory processes to model genre preferences for personalized music rec- ommendations,

    D. Kowald, E. Lex, and M. Schedl, “Utilizing human memory processes to model genre preferences for personalized music rec- ommendations,” in4th Workshop on Transparency and Explain- ability in Adaptive Systems through User Modeling Grounded in Psychological Theory. Association for Computing Machinery (ACM), 2020

    work page 2020

  21. [21]

    M4singer: A multi-style, multi- singer and musical score provided mandarin singing corpus,

    L. Zhang, R. Li, S. Wang, L. Deng, J. Liu, Y . Ren, J. He, R. Huang, J. Zhu, X. Chenet al., “M4singer: A multi-style, multi- singer and musical score provided mandarin singing corpus,”Ad- vances in Neural Information Processing Systems, vol. 35, pp. 6914–6926, 2022

    work page 2022

  22. [22]

    Opencpop: A High-Quality Open Source Chi- nese Popular Song Corpus for Singing V oice Synthesis,

    Y . Wanget al., “Opencpop: A High-Quality Open Source Chi- nese Popular Song Corpus for Singing V oice Synthesis,” inProc. Interspeech, 2022

    work page 2022

  23. [23]

    Singing V oice Scaling-up: An Introduction to ACE- Opencpop and ACE-KiSing,

    J. Shiet al., “Singing V oice Scaling-up: An Introduction to ACE- Opencpop and ACE-KiSing,” inProc. Interspeech, 2024

    work page 2024

  24. [24]

    Tcsinger: Zero-shot singing voice synthesis with style transfer and multi-level style control,

    Y . Zhanget al., “Tcsinger: Zero-shot singing voice synthesis with style transfer and multi-level style control,” inProc. EMNLP, 2024

    work page 2024

  25. [25]

    Tcsinger 2: Customizable multilingual zero-shot singing voice synthesis,

    ——, “Tcsinger 2: Customizable multilingual zero-shot singing voice synthesis,” inProc. ACL, 2025

    work page 2025

  26. [26]

    Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez

    R. Yuan, H. Lin, S. Guo, G. Zhang, J. Pan, Y . Zang, H. Liu, Y . Liang, W. Ma, X. Duet al., “Yue: Scaling open foun- dation models for long-form music generation,”arXiv preprint arXiv:2503.08638, 2025

    work page arXiv 2025

  27. [27]

    DiffRhythm: Blazingly Fast and Embarrassingly Simple End-to-End Full-Length Song Generation with Latent Diffusion

    Z. Ning, H. Chen, Y . Jiang, C. Hao, G. Ma, S. Wang, J. Yao, and L. Xie, “Diffrhythm: Blazingly fast and embarrassingly sim- ple end-to-end full-length song generation with latent diffusion,” arXiv preprint arXiv:2503.01183, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    DiffRhythm+: Controllable and Flexible Full-Length Song Generation with Preference Optimization

    H. Chen, Y . Jiang, G. Ma, C. Hao, S. Wang, J. Yao, Z. Ning, M. Meng, J. Luan, and L. Xie, “Diffrhythm+: Controllable and flexible full-length song generation with preference optimization,” arXiv preprint arXiv:2507.12890, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation

    Z. Liu, S. Ding, Z. Zhang, X. Dong, P. Zhang, Y . Zang, Y . Cao, D. Lin, and J. Wang, “Songgen: A single stage auto- regressive transformer for text-to-song generation,”arXiv preprint arXiv:2502.13128, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    WeaveMuse: An Open Agentic System for Multimodal Music Understanding and Generation

    E. Karystinaios, “Weavemuse: An open agentic system for mul- timodal music understanding and generation,”arXiv preprint arXiv:2509.11183, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Team, “Suno,” https://suno.com/, accessed: 2026.02.01

    S. Team, “Suno,” https://suno.com/, accessed: 2026.02.01

    work page 2026

  32. [32]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  33. [33]

    Mel-band roformer for music source separation,

    J.-C. Wang, W.-T. Lu, and M. Won, “Mel-band roformer for music source separation,” 2023. [Online]. Available: https: //api.semanticscholar.org/CorpusID:263608675

    work page 2023

  34. [34]

    Stars: A unified framework for singing transcrip- tion, alignment, and refined style annotation,

    W. Guo, Y . Zhang, C. Pan, Z. Zhu, R. Li, Z. Chen, W. Xu, F. Wu, and Z. Zhao, “Stars: A unified framework for singing transcrip- tion, alignment, and refined style annotation,” inFindings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 15 081–15 093

    work page 2025

  35. [35]

    Muq: Self-supervised music representation learn- ing with mel residual vector quantization,

    H. Zhu, Y . Zhou, H. Chen, J. Yu, Z. Ma, R. Gu, Y . Luo, W. Tan, and X. Chen, “Muq: Self-supervised music representation learn- ing with mel residual vector quantization,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    work page 2025

  36. [36]

    ESPnet: End-to-End Speech Processing Toolkit

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. E. Y . Soplin, J. Heymann, M. Wiesner, N. Chenet al., “Espnet: End-to-end speech processing toolkit,”arXiv preprint arXiv:1804.00015, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [37]

    V oice evaluation of reason- ing ability: Diagnosing the modality-induced performance gap,

    Y . Lin, Z. Hu, Q. Wang, Y . Liu, H. Zhang, J. Subramanian, N. Vlassis, H. H. Li, and Y . Chen, “V oice evaluation of reason- ing ability: Diagnosing the modality-induced performance gap,” arXiv preprint arXiv:2509.26542, 2025

    work page arXiv 2025

  38. [38]

    Per- formance of chatgpt, gemini and deepseek for non-critical triage support using real-world conversations in emergency department,

    S. Lee, S. Jung, J.-H. Park, H. Cho, S. Moon, and S. Ahn, “Per- formance of chatgpt, gemini and deepseek for non-critical triage support using real-world conversations in emergency department,” BMC Emergency Medicine, vol. 25, no. 1, p. 176, 2025

    work page 2025

  39. [39]

    Emergenttts- eval: Evaluating tts models on complex prosodic, expressiveness, and linguistic challenges using model-as-a-judge,

    R. R. Manku, Y . Tang, X. Shi, M. Li, and A. Smola, “Emergenttts- eval: Evaluating tts models on complex prosodic, expressiveness, and linguistic challenges using model-as-a-judge,”Advances in Neural Information Processing Systems, vol. 38, 2026

    work page 2026

  40. [40]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen et al., “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabil- ities,”arXiv preprint arXiv:2507.06261, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    SingMOS: An extensive Open-Source Singing Voice Dataset for MOS Prediction

    Y . Tang, J. Shi, Y . Wu, and Q. Jin, “Singmos: An extensive open- source singing voice dataset for mos prediction,”arXiv preprint arXiv:2406.10911, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    Singmos-pro: An comprehensive benchmark for singing quality assessment,

    Y . Tang, L. Liu, W. Feng, Y . Zhao, J. Han, Y . Yu, J. Shi, and Q. Jin, “Singmos-pro: An comprehensive benchmark for singing quality assessment,”arXiv preprint arXiv:2510.01812, 2025

    work page arXiv 2025

  43. [43]

    MOS-Bench: Benchmarking Generalization Abilities of Subjective Speech Quality Assessment Models

    W.-C. Huang, E. Cooper, and T. Toda, “Mos-bench: Benchmark- ing generalization abilities of subjective speech quality assess- ment models,”arXiv preprint arXiv:2411.03715, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Versa: A versatile evalu- ation toolkit for speech, audio, and music,

    J. Shi, H.-j. Shim, J. Tian, S. Arora, H. Wu, D. Petermann, J. Q. Yip, Y . Zhang, Y . Tang, W. Zhanget al., “Versa: A versatile evalu- ation toolkit for speech, audio, and music,” inProc. NAACL, 2025

    work page 2025