Pith. sign in

REVIEW 6 major objections 5 minor 40 references

STARS: A Unified Framework for Singing Transcription, Alignment, and Refined Style Annotation

T0 review · 6 major / 5 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 STARS claims that one model can align lyrics, transcribe notes, and annotate vocal style in a single pass, replacing the fragmented toolchains used today.

desk verdict Useful engineering integration, but the reported superiority over baselines is unverified because the comparisons are not matched; send to review to fix the evaluation protocol. read the letter →

arxiv 2507.06670 v1 pith:5S224X76 submitted 2025-07-09 cs.SD eess.AS

Wenxiang Guo , Yu Zhang , Changhao Pan , Zhiyuan Zhu , Ruiqi Li , Zhetao Chen , Wenhao Xu , Fei Wu
show 1 more author
Zhou Zhao
This is my paper
classification cs.SDeess.AS
keywords automaticsingingannotationvoicesynthesisphonemealignmentnotetranscriptionvocaltechniquedetectionstylehierarchicalacousticencodingmultitasklearning
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

STARS is designed to show that one automatic system can annotate a singing recording end to end—aligning lyrics to audio, transcribing notes with pitch and timing, labeling nine vocal techniques at the phoneme level, and classifying global style attributes like emotion and pace—where earlier tools each handled only one of these jobs. The paper reports that this unified model beats the separate alignment, transcription, and technique baselines on their own metrics, and that singing-voice-synthesis models trained on STARS-annotated data sound and respond to style prompts almost as well as models trained on human annotations. The underlying motivation is scalability: if one model can replace fragmented toolchains without cascading errors, building large expressive singing corpora becomes far cheaper. This matters because modern singing voice synthesis is limited less by model architecture than by the scarcity of precisely annotated audio.

What carries the argument

The load-bearing mechanism is the Unified Multi-Level Framework: a shared acoustic backbone, built from a U-Net-style encoder with Conformer blocks and a FreqMOE frequency mixture-of-experts layer, produces frame-level features that are then segment-pooled, vector-quantized, and length-regulated into word, phoneme, note, and sentence representations. Alignment comes from frame-level phoneme logits decoded by Viterbi forced alignment and trained with CTC loss; note boundaries are predicted from fused frame, word, and phoneme features; and pitch plus technique labels are computed by CIF-style attention-weighted pooling over the relevant segments. The same shared features feed a cross-attention global style predictor, so every output conditions on the same acoustic representation rather than on separate model outputs.

What would settle it

Re-annotate a held-out sample of the test songs with two independent expert annotators and measure their inter-annotator agreement on phoneme boundaries and technique labels; if that agreement is close to or below the reported gap between STARS and the baselines, the claimed superiority could reflect label noise rather than genuine transcription and alignment skill.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a single five-level hierarchical encoder—shared across frame, word, phoneme, note, and sentence granularities—can jointly solve lyric alignment, note transcription, technique identification, and global style classification in one forward pass, and that this joint solution is more accurate than specialized single-task systems. In the experiments, STARS reports a Boundary Error Rate of 18.6 versus 20.9 for the strongest forced-alignment baseline and a note-transcription COnPOff of 71.0 versus 70.2, with raw pitch accuracy of 86.7 versus 83.8. Its average technique-prediction F1 and accuracy also exceed the corpus's own technique predictor. The paper further claims that synthesis models trained only on STARS-generated labels reach near-identical perceptual quality and style-control scores to models trained on ground-truth labels.

Load-bearing premise

The whole setup depends on the human-created annotations in the training corpora being accurate and consistent; if those labels contain systematic errors, STARS will faithfully learn and reproduce them, and the reported accuracy numbers would be inflated rather than real.

Editorial extensions

If this is right

  • If the reported results hold, a single STARS forward pass can produce training-ready annotations—phoneme boundaries, note events, techniques, and global style—from raw audio and lyrics.
  • Singing voice synthesis systems can be trained on fully automatic annotations with almost no drop in perceived naturalness or style controllability, lowering dataset production cost.
  • Because the model is trained and evaluated on Chinese and English data through one shared hierarchy, the same toolchain can plausibly be extended to other languages by retraining on in-language corpora.
  • Joint modeling should reduce the cascading errors that arise when alignment, transcription, and style tools are chained independently.

Reading between the lines

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

  • An implication the paper leaves implicit is that the same architecture could serve as a front end for expressive text-to-speech or speech editing, where phoneme boundaries, pitch, and style labels are needed simultaneously.
  • A testable extension would be to evaluate STARS on a sung corpus with independent expert annotation in a third language; if the reported gains depend heavily on the training corpora's label distribution, degradation should appear there first.
  • The observed improvement when mixing predicted labels with half ground truth in synthesis training suggests automatic annotations may act as a regularizer or data augmenter, a hypothesis the paper does not directly test.
  • One could analyze which of the nine techniques are most confused and check whether STARS's errors track perceptual confusability among similar techniques, such as breathy versus weak.
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

6 major / 5 minor

Summary. The paper presents STARS, a unified multi-level neural framework for singing annotation that jointly addresses phoneme-level forced alignment, note transcription and alignment, phone-level vocal technique prediction, and global style attribute prediction. The architecture processes Mel-spectrograms and F0 with a shared CMU encoder (U-Net plus Conformer and FreqMOE), vector quantization at intermediate levels, and Viterbi-based alignment, producing annotations for Chinese and English singing audio. The authors evaluate STARS against MFA and SOFA for alignment, against VOCANO and ROSVOT for note transcription, against GTSinger for technique prediction, and in a downstream singing-voice-synthesis experiment with TCSinger, reporting improved metrics across most tasks.

Significance. If the reported results hold under matched evaluation protocols, STARS is a useful and original contribution: it is, to my knowledge, the first single model that outputs phoneme boundaries, MIDI notes, phone-level techniques, and global style attributes in one forward pass, and the SVS experiment suggests that its annotations are usable in practice. The architecture is well motivated, and the paper demonstrates the framework on both Chinese and English data. However, the current evidence for the central claim of 'superior performance across multiple evaluation dimensions' is weakened by comparison-protocol uncertainties, a missing baseline for global style, and an ablation narrative that contradicts its own table.

major comments (6)
  1. [§4.1.4 and Table 2] The VOCANO and ROSVOT rows in Table 2 are taken from numbers 'reported in ROSVOT' rather than re-run on the same test set as STARS. The table caption does not state the dataset, split, or evaluation criteria for these baselines. Since the headline gains are small (COnPOff 71.0 vs 70.2; RPA 86.7 vs 83.8), a mismatch in test data or in onset/offset tolerance would invalidate the claimed superiority in note transcription. Please re-run the baselines on exactly the STARS test set and report the protocol, or clearly state why the published numbers are directly comparable.
  2. [§4.1.4 and Table 1] The MFA and SOFA results in Table 1 are reported without specifying whether they were evaluated on the same songs, with the same ground-truth boundaries, and with identical tolerance settings as STARS. The 20 ms BER definition is given in §4.1.3, but it is not stated that the baselines were subject to the identical pipeline. Please specify the exact evaluation protocol applied to MFA and SOFA, or re-run them under the same protocol, so that the BER/IOU comparison is meaningful.
  3. [§4.2.3 and Table 4] The ablation narrative is contradicted by the table it refers to. The text states that 'when VQ is omitted for the phone, note, and word levels, the note, technique, and style recognition tasks show improved performance,' but Table 4 shows the w/o VQ row has lower COnPOff (70.7 vs 71.0), lower RPA (86.4 vs 86.7), lower T-F1 (76.3 vs 79.9), lower T-ACC (90.4 vs 91.5), and lower S-ACC (65.3 vs 68.0). Only BER improves (18.3 vs 18.6). This misreporting must be corrected and the conclusions about the vector-quantization component redrawn.
  4. [Tables 1–4] No significance tests, confidence intervals, or multiple-seed variance are reported for any of the objective comparisons. Many of the reported improvements are within one or two points (e.g., COnPOff 71.0 vs 70.2, RPA 86.7 vs 83.8), so without statistical evidence the claim of 'superior performance' is not established. Please add significance testing or variance across at least three training runs, and report the number of test items used for each metric.
  5. [§4.2.2 and Table 3] For global style prediction, Table 3 reports only STARS's accuracy (STY = 68.0) and no baseline is provided. Since the abstract and Section 4.2 claim superior performance across multiple evaluation dimensions, the lack of any comparison for the global-style dimension leaves that part of the claim unsupported. Please add a baseline for emotion, pace, and range prediction, or explicitly restrict the superiority claim to the tasks for which a baseline exists.
  6. [Appendix C.1 and §4.1.1] The 30-hour dataset's ground-truth annotations are described as produced by MFA with manual adjustments, but no inter-annotator agreement, adjustment statistics, or label-quality audit is reported. Because all results in Tables 1–3 are measured against these labels, systematic annotation noise would propagate into both training and evaluation. Please provide at least a small validation sample or a quantitative audit (e.g., percentages of manual boundary corrections) to support the reliability of the ground truth.
minor comments (5)
  1. [§3.7 vs §3.6] Section 3.7 lists Ltech as 'calculated as the CE loss' for technique prediction, while Section 3.6 correctly defines the technique loss as binary cross-entropy (BCE). Please align the notation.
  2. [§4.2.3] There is a broken sentence: 'demonstrating the effectiveness of Finally, experiments on bilingual datasets...' This appears to be a typesetting error and should be corrected.
  3. [Appendix C.3] There is a typo in the sentence 'n contrast, the model performs less effectively on attributes like emotion and vocal range' – 'n' should be 'In'.
  4. [Abstract and §2.2] The dataset name 'VocalSet' is rendered with a space as 'V ocalSet' in multiple places; please fix the formatting.
  5. [Appendix C.3 and Table 8] The final sentence of Appendix C.3 says emotion and vocal range 'fluctuate within different sections' but the preceding text states 'the model performs less effectively on attributes like emotion and vocal range, which may vary across the entire singing performance.' Please reconcile the two descriptions.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: STARS' predictions are trained and evaluated against held-out ground truth; dense self-citations are contextual, not load-bearing.

full rationale

STARS is an empirical system paper. Its derivation chain is: build a hierarchical encoder; train with CE/CTC/BCE losses against GTSinger and newly collected ground-truth labels; then measure BER/IOU, COnPOff/RPA, F1/accuracy on reserved songs. Each predicted quantity (phoneme boundaries, note boundaries/pitch, techniques, global styles) is a model output evaluated against held-out annotations, and no predicted quantity is defined from the fitted parameters that produce it. The technique and alignment labels come from manual annotation and MFA with manual adjustment (Appendix C.1), not from STARS itself. Baselines include external tools (MFA, SOFA, VOCANO) and same-group predictors (ROSVOT, GTSinger); copying ROSVOT's reported numbers is a protocol-matching concern, not circularity, because those numbers are independent measurements outside this paper's fitted values. The many self-citations (ROSVOT, GTSinger, TCSinger, FreqMOE follow-ons) are contextual or methodological, and no load-bearing premise is justified solely by an unverified self-citation. Appendix C.2's 'we employ STARS to exclude silent notes' refers to parsing the model's own predicted note events for evaluation, which is standard practice, not re-fitting ground truth. No equation equates a predicted output to an input, and no fitted parameter is renamed as a prediction. The 'first unified framework' claim is an assertion, not a uniqueness theorem imported from prior work. The score of 2 reflects the density of benign self-citations, not actual circularity.

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

The framework is a trained neural system, so its behavior rests on the quality of the gold-standard annotations, the availability of lyrics, and the architectural hyperparameters listed. No new physical entities are introduced. The evaluation depends on tolerance choices and a small test set.

free parameters (4)
  • VQ codebook size = 128
    Hand-chosen; no ablation over codebook size is reported, so the central model's behavior depends on this choice.
  • FreqMOE expert count = 4
    Hand-chosen; the FreqMOE module splits features into K bands. The paper ablates the module entirely but not the number of experts.
  • Downsampling rate = 16x
    U-Net downsampling factor chosen without sensitivity analysis; it affects the temporal resolution of all boundary predictions.
  • Evaluation tolerances = BER 20ms, note 50ms/50 cents
    These thresholds determine all reported metrics; they are standard in the field but chosen for evaluation, not fitted to the claim.
assumptions (4)
  • domain assumption Ground-truth annotations in GTSinger and the collected 30-hour dataset are accurate and consistent.
    The model is trained and evaluated entirely against these labels; no label-quality audit is performed (Section 4.1.1).
  • domain assumption Lyrics are reliably available, either provided or produced by an ASR system.
    The pipeline needs the phoneme sequence as input for alignment and technique prediction (Section 3.2).
  • domain assumption The 30-song validation/test set is representative of the target distribution.
    All objective results are reported on this small set, selected for containing various techniques and styles without a specified sampling scheme (Section 4.1.1).
  • standard math Viterbi alignment and CTC are valid alignment tools for singing with extended phoneme durations.
    The alignment algorithm relies on standard dynamic programming; the paper does not justify its validity for singing-specific phenomena like long holds (Sections 3.4, Appendix B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of STARS: A Unified Framework for Singing Transcription, Alignment, and Refined Style Annotation." pith.science (2026). https://pith.science/paper/5S224X76

@misc{pith2026250706670,
  author       = {Pith},
  title        = {Pith review of: STARS: A Unified Framework for Singing Transcription, Alignment, and Refined Style Annotation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5S224X76}},
  note         = {Machine review of arXiv:2507.06670}
}
read the original abstract

Recent breakthroughs in singing voice synthesis (SVS) have heightened the demand for high-quality annotated datasets, yet manual annotation remains prohibitively labor-intensive and resource-intensive. Existing automatic singing annotation (ASA) methods, however, primarily tackle isolated aspects of the annotation pipeline. To address this fundamental challenge, we present STARS, which is, to our knowledge, the first unified framework that simultaneously addresses singing transcription, alignment, and refined style annotation. Our framework delivers comprehensive multi-level annotations encompassing: (1) precise phoneme-audio alignment, (2) robust note transcription and temporal localization, (3) expressive vocal technique identification, and (4) global stylistic characterization including emotion and pace. The proposed architecture employs hierarchical acoustic feature processing across frame, word, phoneme, note, and sentence levels. The novel non-autoregressive local acoustic encoders enable structured hierarchical representation learning. Experimental validation confirms the framework's superior performance across multiple evaluation dimensions compared to existing annotation approaches. Furthermore, applications in SVS training demonstrate that models utilizing STARS-annotated data achieve significantly enhanced perceptual naturalness and precise style control. This work not only overcomes critical scalability challenges in the creation of singing datasets but also pioneers new methodologies for controllable singing voice synthesis. Audio samples are available at https://gwx314.github.io/stars-demo/.

Figures

Figures reproduced from arXiv: 2507.06670 by the authors.

Figure 1
Figure 1. STARS vs. Traditional Stepwise Pipeline. Conventional stepwise processing requires sequential execution [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of STARS. (a) The unified multi-level framework, which integrates singing lyric [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages

  1. [1]

    Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. 2023. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747

    arXiv 2023

  2. [2]

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759

    arXiv 2024

  3. [3]

    Linhao Dong and Bo Xu. 2020. Cif: Continuous integrate-and-fire for end-to-end speech recognition. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6079--6083. IEEE

    work page 2020

  4. [4]

    G David Forney. 1973. The viterbi algorithm. Proceedings of the IEEE, 61(3):268--278

    work page 1973

  5. [5]

    Alex Graves, Santiago Fern \'a ndez, Faustino Gomez, and J \"u rgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369--376

    2006

  6. [6]

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. 2020. Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100

    arXiv 2020

  7. [7]

    Wenxiang Guo, Yu Zhang, Changhao Pan, Rongjie Huang, Li Tang, Ruiqi Li, Zhiqing Hong, Yongqi Wang, and Zhou Zhao. 2025. Techsinger: Technique controllable multilingual singing voice synthesis via flow matching. arXiv preprint arXiv:2502.12572

    arXiv 2025

  8. [8]

    Jui-Yang Hsu, Li Su, et al. 2021. Vocano: A note transcription framework for singing voice in polyphonic music

    work page 2021

Show all 40 references
  1. [9]

    Jiawen Huang, Emmanouil Benetos, and Sebastian Ewert. 2022 a . Improving lyrics alignment through joint pitch detection. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 451--455. IEEE

    2022

  2. [10]

    Rongjie Huang, Chenye Cui, Feiyang Chen, Yi Ren, Jinglin Liu, Zhou Zhao, Baoxing Huai, and Zhefeng Wang. 2022 b . Singgan: Generative adversarial network for high-fidelity singing voice generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2525--2535

    2022

  3. [11]

    Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In International Conference on Machine Learning, pages 5530--5540. PMLR

    2021

  4. [12]

    Sungjae Kim, Yewon Kim, Jewoo Jun, and Injung Kim. 2023. Muse-svs: Multi-singer emotional singing voice synthesizer that controls emotional intensity. IEEE/ACM Transactions on Audio, Speech, and Language Processing

    2023

  5. [13]

    Ruiqi Li, Yu Zhang, Yongqi Wang, Zhiqing Hong, Rongjie Huang, and Zhou Zhao. 2024. https://arxiv.org/abs/2405.09940 Robust singing voice transcription serves synthesis . Preprint, arXiv:2405.09940

    2024 arXiv

  6. [14]

    Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. 2022. Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 11020--11028

    2022

  7. [15]

    Peiling Lu, Jie Wu, Jian Luan, Xu Tan, and Li Zhou. 2020. Xiaoicesing: A high-quality and integrated singing voice synthesis system. arXiv preprint arXiv:2006.06261

    2020 arXiv

  8. [16]

    Ilaria Manco, Emmanouil Benetos, Elio Quinton, and György Fazekas. 2021. https://doi.org/10.1109/IJCNN52387.2021.9533461 Muscaps: Generating captions for music audio . In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1--8

    2021

  9. [17]

    Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. 2017. Montreal forced aligner: Trainable text-speech alignment using kaldi. In Interspeech, volume 2017, pages 498--502

    2017

  10. [18]

    Emilio Molina, Ana Maria Barbancho-Perez, Lorenzo Jose Tardon-Garcia, Isabel Barbancho-Perez, et al. 2014. Evaluation framework for automatic singing transcription

    2014

  11. [19]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492--28518. PMLR

    2023

  12. [20]

    Colin Raffel, Brian McFee, Eric J Humphrey, Justin Salamon, Oriol Nieto, Dawen Liang, Daniel PW Ellis, and C Colin Raffel. 2014. Mir\_eval: A transparent implementation of common mir metrics. In ISMIR, volume 10, page 2014

    2014

  13. [21]

    Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2019. Fastspeech: Fast, robust and controllable text to speech. Advances in neural information processing systems, 32

    2019

  14. [22]

    Yi Ren, Xu Tan, Tao Qin, Jian Luan, Zhou Zhao, and Tie-Yan Liu. 2020. Deepsinger: Singing voice synthesis with data mined from the web. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1979--1989

    2020

  15. [23]

    S Resna and Rajeev Rajan. 2023. Multi-voice singing synthesis from lyrics. Circuits, Systems, and Signal Processing, 42(1):307--321

    2023

  16. [24]

    David Snyder, Guoguo Chen, and Daniel Povey. 2015. Musan: A music, speech, and noise corpus. arXiv preprint arXiv:1510.08484

    2015 arXiv

  17. [25]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning. Advances in neural information processing systems, 30

    2017

  18. [26]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  19. [27]

    Jun-You Wang, Chon-In Leong, Yu-Chen Lin, Li Su, and Jyh-Shing Roger Jang. 2023. Adapting pretrained speech model for mandarin lyrics transcription and alignment. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1--8. IEEE

    2023

  20. [28]

    Xianke Wang, Bowen Tian, Weiming Yang, Wei Xu, and Wenqing Cheng. 2022 a . Musicyolo: A vision-based framework for automatic singing transcription. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:229--241

    2022

  21. [29]

    Yongqi Wang, Ruofan Hu, Rongjie Huang, Zhiqing Hong, Ruiqi Li, Wenrui Liu, Fuming You, Tao Jin, and Zhou Zhao. 2024. Prompt-singer: Controllable singing-voice-synthesis with natural language prompt. arXiv preprint arXiv:2403.11780

    2024 arXiv

  22. [30]

    Yu Wang, Xinsheng Wang, Pengcheng Zhu, Jie Wu, Hanzhao Li, Heyang Xue, Yongmao Zhang, Lei Xie, and Mengxiao Bi. 2022 b . Opencpop: A high-quality open source chinese popular song corpus for singing voice synthesis. arXiv preprint arXiv:2201.07429

    2022 arXiv

  23. [31]

    Haojie Wei, Xueke Cao, Tangpeng Dan, and Yueguo Chen. 2023. Rmvpe: A robust model for vocal pitch estimation in polyphonic music. arXiv preprint arXiv:2306.15412

    2023 arXiv

  24. [32]

    Julia Wilkins, Prem Seetharaman, Alison Wahl, and Bryan Pardo. 2018. Vocalset: A singing voice dataset. In ISMIR, pages 468--474

    2018

  25. [33]

    Jie Wu and Jian Luan. 2020. Adversarially trained multi-singer sequence-to-sequence singing synthesizer. arXiv preprint arXiv:2006.10317

    2020 arXiv

  26. [34]

    Siwei Wu, Jinzheng He, Ruibin Yuan, Haojie Wei, Xipin Wei, Chenghua Lin, Jin Xu, and Junyang Lin. 2024. Songtrans: An unified song transcription and alignment method for lyrics and notes. arXiv preprint arXiv:2409.14619

    2024 arXiv

  27. [35]

    Yongmao Zhang, Jian Cong, Heyang Xue, Lei Xie, Pengcheng Zhu, and Mengxiao Bi. 2022. Visinger: Variational inference with adversarial learning for end-to-end singing voice synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (...

    2022

  28. [36]

    Yu Zhang, Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Tao Jin, and Zhou Zhao. 2025 a . Isdrama: Immersive spatial drama generation through multimodal prompting. arXiv preprint arXiv:2504.20630

    2025

  29. [37]

    Yu Zhang, Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Ruiqi Li, Jingyu Lu, Rongjie Huang, Ruiyuan Zhang, Zhiqing Hong, Ziyue Jiang, et al. 2025 b . Versatile framework for song generation with prompt-based control. arXiv preprint arXiv:2504.19062

    2025

  30. [38]

    Yu Zhang, Rongjie Huang, Ruiqi Li, JinZheng He, Yan Xia, Feiyang Chen, Xinyu Duan, Baoxing Huai, and Zhou Zhao. 2024 a . Stylesinger: Style transfer for out-of-domain singing voice synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19...

    2024

  31. [39]

    Yu Zhang, Ziyue Jiang, Ruiqi Li, Changhao Pan, Jinzheng He, Rongjie Huang, Chuxin Wang, and Zhou Zhao. 2024 b . Tcsinger: Zero-shot singing voice synthesis with style transfer and multi-level style control. arXiv preprint arXiv:2409.15977

    2024 arXiv

  32. [40]

    Yu Zhang, Changhao Pan, Wenxiang Guo, Ruiqi Li, Zhiyuan Zhu, Jialei Wang, Wenhao Xu, Jingyu Lu, Zhiqing Hong, Chuxin Wang, et al. 2024 c . Gtsinger: A global multi-technique singing corpus with realistic music scores for all singing tasks. arXiv preprint arXiv:2409.13832

    2024

Pith tools

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