Pith. sign in

REVIEW 4 major objections 4 minor 44 references

Is One Score Enough? Assessing Singing Quality of Songs with Temporal Score Curves

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read SongSQA predicts full-length song singing quality as both an overall score and a temporal segment curve, outperforming clip-level SQA baselines.

desk verdict A real step toward full-song singing quality assessment with temporal score curves, but the empirical claims need baseline details, error bars, and teacher validation before I'd trust the headline numbers. read the letter →

arxiv 2607.16599 v1 pith:UNXGUXC6 submitted 2026-07-18 cs.SD cs.MMeess.AS

Yishan Lv , Jing Luo , Xinyu Yang , Zhizheng Wu This is my paper
classification cs.SDcs.MMeess.AS
keywords singingqualityassessmentfull-lengthsongtemporalscorecurvepseudo-labellearninglearnableembeddingself-attentionaggregationself-supervisedmusicrepresentationMOSprediction
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

Full-length songs cannot be judged fairly by one overall number, the paper argues, because singing quality rises and falls across sections. The authors propose SongSQA, a two-stage framework that first learns to score ten-second segments using pseudo-labels from a pre-trained song-aesthetics teacher model, then aggregates segment features and predicted scores through a learnable song embedding and self-attention to predict the song-level score. The same pipeline produces a temporal score curve that highlights strong and weak passages. On two datasets—one public, one internal expert-annotated—SongSQA beats clip-level SQA systems and simple self-supervised pooling baselines on all four metrics, including up to a 13.95% relative gain in Kendall's tau. If correct, the result gives music AI systems a way to turn a single scalar rating into actionable, time-localized feedback.

What carries the argument

The central mechanism is the combination of pseudo-label supervision with a learnable-song-embedding aggregator. The teacher model supplies segment-level targets from a song-aesthetics model's singing-quality branch; the Segment Score Predictor learns from those targets using dual-path pooling to encode both transient vocal events and overall stability. The Song Quality Aggregator then maps each segment's features and predicted score into a fused embedding, prepends a randomly initialized 'song embedding' (a special token that accumulates song-wide context through self-attention), and uses a two-layer Transformer to let that token attend to all segments. A consistency regularizer (the mean p

What would settle it

Have independent human raters score a random set of individual segments (not whole songs) and compute correlation/agreement between the Segment Score Predictor's segment-level predictions and the human segment-level ratings. If the predictor does not agree with human per-segment judgments beyond chance, the pseudo-label supervision is not meaningful and the temporal curve cannot be trusted.

Watch

Extended reading notes

Core claim

SongSQA models a full song as ordered segments, learns segment scores from teacher pseudo-labels, and aggregates them via a learnable song embedding with self-attention. The first stage (Segment Score Predictor) uses MuQ features, self-attention, and dual-path pooling (max for transient events, mean for stability) to score each 10-second segment, trained on pseudo-labels from a pre-trained song-aesthetics teacher's singing branch. The second stage fuses segment features and predicted scores, prepends a randomly initialized learnable song embedding, and passes the sequence through a two-layer Transformer; the song embedding then drives a final score, with a consistency loss keeping average se

Load-bearing premise

The load-bearing premise is that the pseudo-labels emitted by the pre-trained teacher model are valid segment-level singing-quality judgments; the teacher is never validated against human segment-level ratings before being used as supervision, and the only perceptual check is a five-song, twelve-listener pairwise test.

Editorial extensions

If this is right

  • Full-length song SQA can produce a temporal score curve, so a single overall score is no longer the only output available.
  • Segment-level training can be sourced from a teacher model's pseudo-labels, sidestepping the lack of human segment annotations.
  • The learnable-song-embedding aggregator outperforms mean pooling and clip-level baselines across all reported metrics.
  • The consistency regularizer preserves segment-level semantics while optimizing the song-level objective.
  • A small paired listening test (86.7% agreement) indicates the predicted curve reflects human perception, though at limited scale.

Reading between the lines

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

  • The teacher model's segment-level judgments are the de facto ground truth for Stage I; if that teacher is biased or noisy at segment granularity, the entire curve inherits the bias. An editor would expect the performance ceiling to be set by the teacher's segment-level validity, not by the aggregator.
  • The temporal curve could be repurposed as a reward or critique signal for singing-voice synthesis and conversion systems, letting them optimize specific weak passages instead of a single scalar—an application the paper mentions only in passing.
  • The reported 13.95% KTAU gain is relative; absolute KTAU values (≈0.47 on the public dataset, ≈0.71 on the internal one) leave substantial room for improvement, so the claim should be read as 'better than baselines', not 'near-perfect assessment'.
  • A direct human validation of per-segment scores (not just pairwise comparison) would strengthen the curve's credibility; the paper's five-song, twelve-listener study is suggestive but not decisive.
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

4 major / 4 minor

Summary. The paper proposes SongSQA, a two-stage framework for singing quality assessment (SQA) of full-length songs. Stage I trains a Segment Score Predictor on 10-second audio segments using pseudo-labels generated by a pre-trained teacher model (reference [22]); Stage II aggregates the resulting segment features and scores through a Transformer with a learnable song embedding to predict an overall song-level quality score while also producing a temporal segment-level quality curve. The method is evaluated on the public Lyra-SA dataset and a proprietary Internal Dataset against three clip-level SQA baselines (UTMOSv2, PS-SQA, RAMP+) and three SSL-based baselines (wav2vec 2.0, MERT, MuQ). The paper reports consistent improvements on MSE, LCC, SRCC, and KTAU, with a headline 13.95% relative KTAU improvement over the strongest baseline on the Internal Dataset.

Significance. If the reported results are reproducible, the paper addresses a real gap: existing SQA methods are designed for short clips, while full-song evaluation requires modeling quality variation over time. The proposed two-stage design, especially the use of teacher-generated segment pseudo-labels and a learnable song embedding with self-attention, is motivated and the ablation study targets the key components. The use of the public Lyra-SA dataset and the provision of demo audio excerpts are strengths. However, the empirical evidence is currently weakened by several unresolved methodology questions: the possibility of song/singer leakage in the data split, the unspecified adaptation of clip-level baselines to full songs, the absence of variability estimates, and the lack of validation of the teacher pseudo-labels. These issues are addressable but need to be fixed before the central claims can be accepted.

major comments (4)
  1. [§4.1 (Datasets and Implementation Details)] The paper states only that the data are 'randomly split into training, validation, and test sets with a ratio of 8:1:1.' This is ambiguous and potentially problematic for both datasets. Lyra-SA consists of 10 songs × 100 covers; a random recording-level split will place different covers of the same song in both training and test. The Internal Dataset contains 2,035 recordings from 1,284 songs and 1,250 singers, so the same song or the same singer can appear on both sides of the split. Because SongSQA processes ordered segment sequences from full songs, a model could exploit song identity or singer identity to predict an average quality for that song/singer, inflating all correlations in Table 1. Please clarify whether the splits are at the song/singer level, and report results under strictly disjoint-song and disjoint-singer splits; if such splits were already used, the text should say s
  2. [§4.1 (Baseline Models) and Table 1] The adaptation of the three clip-level baselines (UTMOSv2, PS-SQA, RAMP+) to full-length songs is never described. The text says 'all baselines are retrained on our training sets under identical experimental settings,' but only the SSL-based baselines are explicitly adapted (mean pooling over output embeddings followed by a linear head). UTMOSv2, PS-SQA, and RAMP+ are designed for short clips; full songs are several minutes long. Were they fed entire songs (which may violate their input-length assumptions), or segmented with some pooling strategy? Without specifying segment length, overlap, and score aggregation for these baselines, the results in Table 1 are not reproducible, and the conclusion in §4.2 that these methods 'remain clearly limited when applied to full-length song SQA' is not verifiable. Please document the exact inference protocol and, if needed, rerun them with a well-spe
  3. [Table 1 and Table 2] All numerical results are reported only as arithmetic means over five random splits, with no standard deviations, confidence intervals, or significance tests. Several headline differences are small: for example, in Table 2, SongSQA vs 'w/o Scores' on Lyra-SA has MSE 114.01 vs 115.32 but LCC 0.6237 vs 0.6277 (the ablation is better on LCC); on the Internal Dataset, 'w/o SongEmb & TE' has lower MSE than the full model (35.71 vs 37.56). The claim of 'consistent' improvement therefore rests on unquantified noise. Please report per-split results with variability measures and, where appropriate, paired significance tests (e.g., Wilcoxon signed-rank across the five splits).
  4. [§3.2 (Eqs. 7–8) and §4.3] Stage I is supervised entirely by pseudo-labels from teacher model [22], whose 'Singing' dimension outputs are used as segment-level ground truth. No evidence is presented that these pseudo-labels agree with human segment-level quality judgments; [22] is by overlapping authors and is not independently validated here. The only perceptual check is the 5-song/12-listener pairwise study in §4.5, which tests the final model's curve, not the teacher. Since any systematic bias in the teacher is inherited by the Segment Score Predictor and propagates to the aggregator, the claim that SongSQA learns 'precise vocal details' (Section 1) is not yet supported. Please include a validation of the teacher against human segment ratings (or a held-out annotation subset), or discuss the limitations explicitly.
minor comments (4)
  1. [§4.1 (Implementation Details)] Batch size, number of epochs, learning rate schedule, and early stopping criteria are not reported. The silence-filtering threshold and the maximum sequence length / padding strategy for the aggregator are also unspecified. These details are needed for reproducibility.
  2. [Table 2] In the rendered text, several numbers run together (e.g., 'w/o Scores 115.320.62770.62840.4681'), making the table difficult to read. Please fix the formatting so each column is clearly separated.
  3. [§4.5] The subjective evaluation is small (5 songs, 12 listeners) and selects the two segments based on the model's own predictions, which can bias the comparison toward extreme pairs. Please report the criterion for selecting songs and segments and discuss the limitations of this protocol. An archived demo link or supplementary audio would also help.
  4. [§4.3 (Teacher Models)] The teacher model [22] and the 'Gemini-3-Pro' system used in the ablation are not described in terms of availability, licensing, or input requirements. Please clarify how these teachers were accessed and whether their checkpoints are publicly released.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation reduces to its inputs; the final song-level evaluation is anchored to external human labels, but the teacher pseudo-labeling step is a self-cited component with limited external validation.

full rationale

I traced the claimed derivation chain. The final song-level prediction is trained with external human labels (Lyra-SA and expert-annotated Internal Dataset), so the headline result in Table 1 is not equivalent to the model's own inputs. Stage I uses pseudo-labels from a teacher model [22] via Eq. (7) and Eq. (8), but this is an intermediate imitation objective, not a claim that the teacher generates ground truth. The consistency loss Eq. (16) ties only the average segment score to the song-level label as a regularizer; the temporal curve's shape remains unconstrained and is checked by the small pairwise subjective study in Section 4.5. The teacher model [22] is prior work by overlapping authors and is not independently validated for segment-level singing quality; that is an evidence and reproducibility weakness, not a by-construction equivalence. Similarly, the paper does not specify how clip-level baselines such as UTMOSv2, PS-SQA, and RAMP+ were adapted to full-length songs, which is a comparability concern rather than circularity. No step in the derivation was found to reduce to its own inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The framework rests on the validity of teacher-generated segment pseudo-labels, the decomposability of singing quality into segment-level scores, and the quality of MuQ representations. No segment-level human annotations exist to verify the first two. The only explicitly hand-set hyperparameter is the consistency loss weight; other choices (segment length, transformer depth) are manual but standard.

free parameters (2)
  • Consistency loss weight lambda = 0.1
    Balances L_song and L_cons in Eq. (18); chosen by hand without reported sensitivity analysis.
  • Segment length / overlap = 10 s / 5 s
    Manual preprocessing choice that determines temporal curve granularity and the number of pseudo-label training segments; no ablation is reported.
assumptions (3)
  • domain assumption Pseudo-labels T(v_i) from the teacher model are valid segment-level singing-quality supervision.
    Stage I is trained entirely against these labels (Eq. 8), but the teacher is from overlapping authors and is not validated against human segment ratings.
  • domain assumption Overall singing quality decomposes into isolated 10-second segment scores plus attention-based aggregation that mimics human perception.
    Motivates the architecture in Sec. 3.3; supported only by citations [11,25,42], not by measured segment-level ground truth.
  • domain assumption MuQ SSL representations are informative for perceptual singing quality.
    MuQ is used as the feature backbone in Eq. (1); its adequacy for singing quality is assumed from prior SSL literature rather than demonstrated here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is One Score Enough? Assessing Singing Quality of Songs with Temporal Score Curves." pith.science (2026). https://pith.science/paper/UNXGUXC6

@misc{pith2026260716599,
  author       = {Pith},
  title        = {Pith review of: Is One Score Enough? Assessing Singing Quality of Songs with Temporal Score Curves},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNXGUXC6}},
  note         = {Machine review of arXiv:2607.16599}
}
read the original abstract

Singing Quality Assessment (SQA) has become increasingly important for practical multimedia applications and Music AI systems, yet existing studies predominantly focus on short singing clips and remain insufficient for full-length songs. Unlike clip-level assessment, full-length song SQA requires modeling how singing quality varies across different audio segments and how these local variations influence the overall evaluation of vocal performance. Moreover, the scarcity of segment-level annotations makes effective supervision challenging, as directly assigning a single overall score label to every segment tends to treat different segment qualities as equivalent. To address these challenges, we propose SongSQA, a two-stage framework for full-length song SQA. In the first stage, a Segment Score Predictor is trained with pseudo labels generated by a pre-trained teacher model, enabling segment-level singing quality prediction without requiring manual segment annotations. In the second stage, a Song Quality Aggregator integrates segment features and predicted segment scores into unified segment embeddings, and employs a learnable song embedding together with self-attention to capture the connection between segment-level vocal performance and overall song quality. In this way, SongSQA dynamically aggregates critical quality cues across the song to produce a holistic quality prediction, while also generating a temporal segment-level quality curve. Experimental results demonstrate the effectiveness of SongSQA for full-length song SQA, achieving up to a 13.95% relative improvement in KTAU over the strongest baseline, while consistently improving other evaluation metrics across all datasets.

Figures

Figures reproduced from arXiv: 2607.16599 by the authors.

Figure 1
Figure 1. Temporal score curve of SongSQA on a full-length [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison between conventional clip-level SQA [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Model Architecture 3.3 Song Quality Aggregator Building upon the Segment Score Predictor, the second stage is designed to model the intrinsic connection between segment-level vocal performance and overall song quality. Rather than relying on simple averaging, it aggregates the quality variations across different audio segments within a full-length song and produces a final song-level quality prediction. Given a full… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Predicted temporal score curve generated by [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 5 linked inside Pith

  1. [22]

    Yishan Lv, Jing Luo, Boyuan Ju, Yang Zhang, Xinda Wu, Bo Yuan, and Xinyu Yang

  2. [1]

    Kaito Baba, Wataru Nakata, Yuki Saito, and Hiroshi Saruwatari. 2024. The T05 System for the voicemos challenge 2024: Transfer Learning from Deep Image Classifier to Naturalness MOS Prediction of High-Quality Synthetic Speech. In IEEE Spoken Language Technology Workshop, SLT 2024, Macao, December 2-5,

    2024

  3. [2]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Represen- tations. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., 12449–12460

    2020

  4. [3]

    Peng Bai, Yue Zhou, Ke Gu, Meizhen Zheng, Linshujie Zheng, Yidong Chen, and Xiaodong Shi. 2026. Reference-free singing voice MOS prediction via multi- feature fusion, with integrated feature analysis.Applied Acoustics241 (2026), 110960

    2026

  5. [4]

    Chuan Cao, Ming Li, Xiao Wu, Hongbin Suo, Jian Liu, and Yonghong Yan. 2009. Automatic Singing Performance Evaluation for Untrained Singers.IEICE Trans- actions on Information and SystemsE92-D, 8 (August 2009), 1596–1600

    2009

  6. [5]

    Ping-Chen Chan, Po-Wei Chen, and Von-Wun Soo. 2024. Improve Singing Qual- ity Prediction Using Self-supervised Transfer Learning and Human Perception Feedback. InProceedings of the 5th ACM International Conference on Multimedia in Asia(Tainan, Taiwan)(MMAsia ’23). Association for Computing Machinery, New York, NY, USA, Article 69, 7 pages

    2024

  7. [6]

    Erica Cooper, Wen-Chin Huang, Tomoki Toda, and Junichi Yamagishi. 2022. Generalization Ability of MOS Prediction Networks. InICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 8442–8446

    2022

  8. [7]

    Dannenberg, and Zeyu Jin

    Shuqi Dai, Yunyun Wang, Roger B. Dannenberg, and Zeyu Jin. 2025. Everyone- Can-Sing: Zero-Shot Singing Voice Synthesis and Conversion with Speech Ref- erence. InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 1–5

    2025

Show all 44 references
  1. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human...

    2019

  2. [9]

    Wenxiang Guo, Yu Zhang, Changhao Pan, Rongjie Huang, Li Tang, Ruiqi Li, Zhiqing Hong, Yongqi Wang, and Zhou Zhao. 2025. TechSinger: Technique Con- trollable Multilingual Singing Voice Synthesis via Flow Matching. InProceedings of the AAAI Conference on Artificial Intelligence,...

    2025

  3. [10]

    Chitralekha Gupta, Haizhou Li, and Ye Wang. 2017. Perceptual Evaluation of Singing Quality. In2017 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 577–586

    2017

  4. [11]

    Chitralekha Gupta, Haizhou Li, and Ye Wang. 2018. A Technical Framework for Automatic Perceptual Evaluation of Singing Quality.APSIPA Transactions on Signal and Information Processing7 (2018), e10. Lv et al

    2018

  5. [12]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self- Supervised Speech Representation Learning by Masked Prediction of Hidden Units.IEEE/ACM Transactions on Audio, Speech, and Language Processin...

    2021

  6. [13]

    Lin Huang, Chitralekha Gupta, and Haizhou Li. 2020. Spectral Features and Pitch Histogram for Automatic Singing Quality Evaluation with CRNN. In2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). 492–499

    2020

  7. [14]

    Zezario, Tomoki Toda, Hsin-Min Wang, Junichi Yamagishi, and Yu Tsao

    Wen-Chin Huang, Szu-Wei Fu, Erica Cooper, Ryandhimas E. Zezario, Tomoki Toda, Hsin-Min Wang, Junichi Yamagishi, and Yu Tsao. 2024. The Voicemos Challenge 2024: Beyond Speech Quality Prediction. In2024 IEEE Spoken Language Technology Workshop (SLT). 803–810

    2024

  8. [15]

    Wen-Chin Huang, Hui Wang, Cheng Liu, Yi-Chiao Wu, Andros Tjandra, Wei- Ning Hsu, Erica Cooper, Yong Qin, and Tomoki Toda. 2025. The AudioMOS Challenge 2025.arXiv preprint arXiv:2509.01336(2025)

    2025 arXiv

  9. [16]

    Xin Jin, Wu Zhou, Jinyu Wang, Duo XU, and Yongsen Zheng. 2023. An Order- Complexity Aesthetic Assessment Model for Aesthetic-aware Music Recommen- dation. InProceedings of the 31st ACM International Conference on Multimedia (Ottawa ON, Canada)(MM ’23). Association for Computin...

    2023

  10. [17]

    Minseok Kim, Woosung Choi, Jaehwa Chung, Daewon Lee, and Soonyoung Jung

  11. [18]

    Partha Lal. 2006. A Comparison of Singing Evaluation Algorithms. InNinth International Conference on Spoken Language Processing, INTERSPEECH-ICSLP 2006, Pittsburgh, PA, USA, September 17-21, 2006. ISCA

    2006

  12. [19]

    Jinhu Li, Chitralekha Gupta, and Haizhou Li. 2021. Training Explainable Singing Quality Assessment Network with Augmented Data. In2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). 904–911

    2021

  13. [20]

    Dannenberg, Ruibo Liu, Wenhu Chen, Gus Xia, Yemin Shi, Wenhao Huang, Zili Wang, Yike Guo, and Jie Fu

    Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Cheng- hao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, Norbert Gyenge, Roger B. Dannenberg, Ruibo Liu, Wenhu Chen, Gus Xia, Yemin Shi, Wenhao Huang, Zili Wang, Yike Guo, and Jie Fu. 2024. MERT: Acou...

    2024

  14. [21]

    Cheng Liu, Hui Wang, Jinghua Zhao, Shiwan Zhao, Hui Bu, Xin Xu, Jiaming Zhou, Haoqin Sun, and Yong Qin. 2025. MusicEval: A Generative Music Dataset with Expert Ratings for Automatic Text-to-Music Evaluation. InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Spee...

    2025

  15. [23]

    Emilio Molina, Isabel Barbancho, Emilia Gómez, Ana Maria Barbancho, and Lorenzo J Tardón. 2013. Fundamental Frequency Alignment vs. Note-Based Melodic Similarity for Singing Voice Assessment. In2013 IEEE international conference on acoustics, speech and signal processing. IEEE...

    2013

  16. [24]

    Tomoyasu Nakano, Masataka Goto, and Yuzuru Hiraga. 2006. Subjective Eval- uation of Common Singing Skills Using the Rank Ordering Method. InNinth International Conference on Music Perception and Cognition. Citeseer, 1507–1512

    2006

  17. [25]

    Alexander Rozin, Paul Rozin, and Emily Goldberg. 2004. The Feeling of Music Past: How Listeners Remember Musical Affect.Music Perception22, 1 (09 2004), 15–39

    2004

  18. [26]

    Binzhu Sha, Xu Li, Zhiyong Wu, Ying Shan, and Helen Meng. 2024. Neural Concatenative Singing Voice Conversion: Rethinking Concatenation-Based Ap- proach for One-Shot Singing Voice Conversion. InICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Pr...

    2024

  19. [27]

    Yu-Fei Shi, Yang Ai, Ye-Xin Lu, Hui-Peng Du, and Zhen-Hua Ling. 2024. Pitch- and-Spectrum-Aware Singing Quality Assessment with Bias Correction and Model Fusion. InIEEE Spoken Language Technology Workshop, SLT 2024, Macao, December 2-5, 2024. IEEE, 811–817

    2024

  20. [28]

    Johan Sundberg. 2001. Level and Center Frequency of the Singer’s Formant. Journal of Voice15, 2 (2001), 176–186

    2001

  21. [29]

    Yuxun Tang, Lan Liu, Wenhao Feng, Yiwen Zhao, Jionghao Han, Yifeng Yu, Jiatong Shi, and Qin Jin. 2026. SingMOS-Pro: An Comprehensive Benchmark for Singing Quality Assessment. InICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 1–5

    2026

  22. [30]

    Yuxun Tang, Jiatong Shi, Yuning Wu, and Qin Jin. 2024. SingMOS: An Exten- sive Open-Source Singing Voice Dataset for MOS Prediction.arXiv preprint arXiv:2406.10911(2024)

    2024 arXiv

  23. [31]

    Andros Tjandra, Yi-Chiao Wu, Baishan Guo, John Hoffman, Brian Ellis, Apoorv Vyas, Bowen Shi, Sanyuan Chen, Matt Le, Nick Zacharov, et al . 2025. Meta Audiobox Aesthetics: Unified Automatic Quality Assessment for Speech, Music, and Sound.arXiv preprint arXiv:2502.05139(2025)

    2025 arXiv

  24. [32]

    Wei-Ho Tsai and Hsin-Chieh Lee. 2012. Automatic Evaluation of Karaoke Singing Based on Pitch, Volume, and Rhythm Features.IEEE Transactions on Audio, Speech, and Language Processing20, 4 (2012), 1233–1243

    2012

  25. [33]

    Hui Wang, Shiwan Zhao, Xiguang Zheng, Jiaming Zhou, Xuechen Wang, and Yong Qin. 2025. RAMP+: Retrieval-Augmented MOS Prediction With Prior Knowledge Integration.IEEE Transactions on Audio, Speech and Language Pro- cessing33 (2025), 1520–1534

    2025

  26. [34]

    Ziqian Wang, Xianjun Xia, Chuanzeng Huang, and Lei Xie. 2026. S 2Voice: Style- Aware Autoregressive Modeling with Enhanced Conditioning for Singing Style Conversion. InICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 1–5

    2026

  27. [35]

    Zihao Wang, Ruibin Yuan, Ziqi Geng, Hengjia Li, Xingwei Qu, Xinyi Li, Songye Chen, Haoying Fu, Roger B Dannenberg, and Kejun Zhang. 2025. Singing Timbre Popularity Assessment Based on Multimodal Large Foundation Model. InProceedings of the 33rd ACM International Conference on ...

    2025

  28. [36]

    Joel Wapnick and Elizabeth Ekholm. 1997. Expert Consensus in Solo Voice Performance Evaluation.Journal of Voice11, 4 (1997), 429–436

    1997

  29. [37]

    Jixun Yao, Guobin Ma, Huixin Xue, Huakang Chen, Chunbo Hao, Yuepeng Jiang, Haohe Liu, Ruibin Yuan, Jin Xu, Wei Xue, et al. 2025. Songeval: A Benchmark Dataset for Song Aesthetics Evaluation.arXiv preprint arXiv:2505.10793(2025)

    2025 arXiv

  30. [38]

    Ning Zhang, Tao Jiang, Feng Deng, and Yan Li. 2019. Automatic Singing Evalua- tion without Reference Melody Using Bi-dense Neural Network. InICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 466–470

    2019

  31. [39]

    Xueyao Zhang, Junan Zhang, Yuancheng Wang, Chaoren Wang, Yuanzhe Chen, Dongya Jia, Zhuo Chen, and Zhizheng Wu. 2026. Vevo2: A Unified and Control- lable Framework for Speech and Singing Voice Generation.IEEE Transactions on Audio, Speech and Language Processing(2026), 1–17

    2026

  32. [40]

    Yu Zhang, Rongjie Huang, Ruiqi Li, JinZheng He, Yan Xia, Feiyang Chen, Xinyu Duan, Baoxing Huai, and Zhou Zhao. 2024. StyleSinger: Style Transfer for Out- of-Domain Singing Voice Synthesis. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 19597–19605

    2024

  33. [41]

    Haina Zhu, Yizhi Zhou, Hangting Chen, Jianwei Yu, Ziyang Ma, Rongzhi Gu, Yi Luo, Wei Tan, and Xie Chen. 2025. MuQ: Self-Supervised Music Representation Learning With Mel Residual Vector Quantization.IEEE Transactions on Audio, Speech and Language Processing33 (2025), 3653–3664

    2025

  34. [42]

    Nathan Zoanetti and Helen Champion. 2016. Mitigating the Halo Effect: Manag- ing the Wow Factor in Music Performance Assessments.Journal of Research in Music Performance(2016), 36–52

    2016

  35. [2021]

    arXiv preprint arXiv:2111.12203(2021)

    KUIELab-MDX-Net: A Two-Stream Neural Network for Music Demixing. arXiv preprint arXiv:2111.12203(2021)

    2021 arXiv

  36. [2026]

    InProceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR)

    Song Aesthetics Evaluation with Multi-Stem Attention and Hierarchical Uncertainty Modeling. InProceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR). 1–8

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.