Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

TCSinger 2: Customizable Multilingual Zero-shot Singing Voice Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TCSinger 2 claims a single multilingual model can synthesize unseen singers' voices with style drawn from singing, speech, or natural-language text prompts, and outperforms earlier systems on quality, similarity, and controllability.

desk verdict A solid zero-shot SVS system with a credible architecture, but its headline claim rests on underspecified baseline modifications that a referee should force the authors to document. read the letter →

arxiv 2505.14910 v3 pith:IXJWZB2F submitted 2025-05-20 eess.AS cs.CLcs.SD

Yu Zhang , Wenxiang Guo , Changhao Pan , Dongyu Yao , Zhiyuan Zhu , Ziyue Jiang , Yuhan Wang , Tao Jin
show 1 more author
Zhou Zhao
This is my paper · ORCID
classification eess.AScs.CLcs.SD
keywords singingvoicesynthesiszero-shotmultilingualstyletransfercontrolflowmatchingmixtureofexpertscontrastivelearning
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

The paper introduces TCSinger 2, a multilingual zero-shot singing voice synthesis model that generates new singers' voices from lyrics and musical notation without retraining on the target singer. The model's central goal is to make style customizable: a short singing or speech prompt transfers timbre, technique, emotion, accent, and pronunciation, while a natural-language text prompt controls global and segment-level styles such as vocal range, bel canto, happy, mixed voice, or falsetto. The paper argues that three components—blurred boundary content encoding, contrastive audio-prompt alignment, and a flow-matching transformer with language- and style-specific experts and F0 supervision—together remove the usual dependence on exact phoneme and note boundary annotations. The reported experiments claim superior subjective and objective scores over previous singing and speech synthesis baselines across zero-shot style transfer, cross-lingual transfer, multi-level text control, and speech-to-singing tasks. If the claim holds, custom-quality singing voices become producible from a few seconds of reference audio and a text description, across nine languages.

What carries the argument

The load-bearing mechanism is the Blurred Boundary Content (BBC) Encoder: it predicts phoneme durations, expands content embeddings to frame level, then masks $m=8$ tokens at phoneme and note boundaries so the model must learn smooth implicit alignments instead of relying on exact annotations. Around it sit the Custom Audio Encoder, which uses contrastive learning to map singing, speech, and text prompts into one aligned style representation, and the Flow-based Custom Transformer, which uses flow matching with two expert groups (language and style) and F0 prediction to turn content plus style embeddings into a mel-spectrogram. The masking is what makes the system work on noisy, automatically aligned multilingual data; the contrastive alignment is what lets one representation serve all prompt types; the flow-based transformer with Cus-MOE is what carries generation quality and fine-grained style modeling.

What would settle it

Run the released code with the same extended set plus the open corpora, reimplement the two strongest baselines with note encoders of matched parameter count and identical training budget, and compare on the same 30 unseen singers using the same MOS protocol and WavLM cosine similarity; if either adapted baseline matches or exceeds TCSinger 2's MOS-Q, MOS-S, and FFE or Cos numbers, the paper's central claim fails. Alternatively, feed the model deliberately shifted phoneme and note alignments and measure pronunciation accuracy: if quality collapses with small alignment errors, the claimed robustness of the BBC Encoder is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single architecture can handle four related tasks: zero-shot style transfer from unseen singing prompts, cross-lingual style transfer, multi-level style control from textual prompts, and speech-to-singing style transfer. The central design idea is that deliberately blurring phoneme and note boundaries—rather than refining the alignments—makes the model more robust and produces smoother transitions. The Blurred Boundary Content Encoder predicts duration, expands content embeddings to frame level, and randomly masks tokens at boundaries ($m=8$), letting the transformer's self-attention refine implicit alignments. The Custom Audio Encoder aligns singing, speech, and textual prompt embeddings by contrastive learning in a shared style space, and the Flow-based Custom Transformer uses language- and style-routed mixture-of-experts plus F0 supervision to generate mel-spectrograms by flow matching. The paper reports that the full model outperforms all compared baselines on MOS-Q, MOS-S, MOS-C, cosine similarity, and F0 frame error, and that each module's removal lowers comparative MOS scores in ablations.

Load-bearing premise

The central superiority claim rests on the fairness of the baseline adaptations: the speech TTS baselines were given a note encoder and retrained on the same multilingual data, but the paper does not report the architecture or training details of those modifications, so the comparison could be unfair if those baselines were not given equivalent capacity or tuning.

Editorial extensions

If this is right

  • Zero-shot custom voices: a user can generate an unseen singer's voice from lyrics, musical notation, and a few seconds of singing or speech prompt, in all nine supported languages.
  • Text-only style direction: global timbre (gender, vocal range), singing method, emotion, and word-level techniques (falsetto, vibrato, breathy, mixed voice) can be set by a natural-language prompt without any reference audio.
  • Speech-to-singing works: a person who cannot sing can use a speech recording as the prompt and get a singing voice that preserves their timbre and accent.
  • Data-hunger relief: because boundary annotations are deliberately blurred, the model can train on roughly aligned multilingual data and still produce natural transitions.
  • Fast generation: flow matching with classifier-free guidance lets inference run with 25 ODE steps instead of the 1,000 training timesteps, with the paper reporting better quality and controllability at CFG scale 3.

Reading between the lines

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

  • The boundary-masking idea is portable: any annotation-hungry speech or singing model could adopt it to train on automatically aligned data, and the robustness claim would be testable by degrading alignment quality and measuring pronunciation or pitch accuracy.
  • Because singing, speech, and text prompts are contrastively aligned into one latent space, the same encoder could support style retrieval or style interpolation between prompts without retraining; the paper does not test this.
  • The dense-to-sparse routing with per-language and per-style experts suggests a scaling path: adding languages or styles could mean adding experts rather than retraining the whole model, though the paper only studies up to four experts per group.
  • The paper acknowledges the manual style-label bottleneck in its limitations section; replacing it with automatic labeling is the natural extension, but unproven here.
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 / 5 minor

Summary. The paper introduces TCSinger 2, a multilingual zero-shot singing voice synthesis system that supports style transfer from singing or speech prompts and style control from natural-language textual prompts. The system combines three main components: a Blurred Boundary Content (BBC) Encoder that predicts durations and masks phoneme/note boundaries, a Custom Audio Encoder that learns aligned representations across singing, speech, and text via contrastive learning, and a Flow-based Custom Transformer with Cus-MOE and F0 supervision. The authors report subjective and objective evaluations showing TCSinger 2 ahead of StyleTTS 2, CosyVoice, VISinger 2, and TCSinger on style transfer, multi-level style control, and speech-to-singing tasks, together with ablations for each module. The empirical claims are the core of the paper; no formal derivations or theoretical guarantees are offered.

Significance. If the comparative results are reliable, TCSinger 2 would be a meaningful step toward a single model that handles multilingual zero-shot singing synthesis with flexible, multi-level style control. The architecture is plausible, the ablation study supports the utility of each proposed module, and the paper includes limitations and ethics statements plus links to a demo and code. The main gap is evidential: the headline claim that TCSinger 2 outperforms all baselines in both subjective and objective metrics rests on baseline adaptations and hyperparameter choices that are not sufficiently documented, so the magnitude and fairness of the reported advantages are currently hard to assess.

major comments (4)
  1. [§4.1 'Baseline Models'; Appendix D] The central comparative claim — that TCSinger 2 outperforms StyleTTS 2 and CosyVoice on singing tasks — is not verifiable from the manuscript because the modifications to those baselines are underspecified. Section 4.1 states only that the models were 'enhanced with a note encoder to process musical notations' and that a cross-attention model was added for textual prompts; Appendix D lists the official repositories but gives no architecture, fusion point, conditioning granularity, training schedule, or hyperparameters for these additions. Since every main table (Tables 1–3) reports TCSinger 2 ahead of these adapted baselines, the margins could reflect handicapped baselines rather than genuine improvements. Please provide a complete specification of the baseline adaptations and, ideally, release the modified code and training recipes so the comparison can be reproduced.
  2. [§4.1 'Implementation Details'; Appendices E.1–E.2, Tables 6–7] Hyperparameters that directly affect the reported results — the CFG scale γ and the number of experts per Cus-MOE group — are selected using CMOS evaluations relative to the final configuration (γ=3, four experts) in Tables 6 and 7, yet the manuscript never states that these selections were made on a validation split. If the same test items are used for hyperparameter selection and for the main comparisons, the reported margins in Tables 1–3 are optimistically biased. Please specify the validation/test split used for hyperparameter tuning and confirm that the main evaluation set was not used for model selection.
  3. [Tables 1–3; §4.2] The objective metrics supporting the abstract's claim (FFE and Cos) are reported as point estimates without error bars or significance tests, and several margins are small (for example, FFE 0.24 vs 0.28 in Table 3, Cos 0.89 vs 0.82). Since the headline claim explicitly includes objective metrics, please report variance (e.g., bootstrap confidence intervals or per-sample standard errors) or statistical tests over the test set to show that the objective differences are not within noise.
  4. [Appendix B, Table 5] The newly collected 'Extended' subset is load-bearing for the multilingual and cross-lingual claims, but it is described only as '50 hours of clean singing voices' with five hours of speech. The number of singers, language distribution, recording conditions, annotation procedure, and how this subset is partitioned from the 30 unseen singers are not given. Without this information, the reader cannot assess data leakage or the representativeness of the multilingual test set. Please supplement this description.
minor comments (5)
  1. [§4.3, Table 4] The sentence about removing CFG is internally contradictory: 'we find that CMOS-Q, CMOS-S, and CMOS-C decrease significantly, while CMOS-Q only shows a slight decline' names CMOS-Q twice. The 'w/o CFG' row shows CMOS-Q -0.26/-0.25, which is not a slight decline; please clarify which metric is meant.
  2. [§3.3, Eq. (1)] The contrastive loss notation is malformed: the subscripts in L_{p_i_si,p_i_sp} and the total-loss expression L_{pisp,pte} + L_{psi,pte} do not match the definitions in the equation, and the displayed formula appears to use the same positive pair twice without covering the three described contrast types. Please rewrite with consistent superscripts and define every term.
  3. [§3.3] It is unclear how the textual encoder 'combines music scores and textual prompts': the text says cross-attention is used, but Section 3.1 describes the textual prompt as being encoded into z_pt without explaining whether music scores enter the textual branch and how. Please clarify the role of music scores in this path.
  4. [Appendix B, Table 5] The dataset table calls the PopBuTFy subset 'BuTFy', while Section 4.1 and the reference list use 'PopBuTFy'; please make the naming consistent throughout.
  5. [§4.1] The sentence 'All languages include Chinese, English, French, Spanish, German, Italian, Japanese, Korean, and Russian' is grammatically incomplete; please state which items — training data, test data, or all datasets — contain these languages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical system comparisons, not derivations that reduce to their own inputs.

full rationale

The central claims of TCSinger 2 are empirical: the model is trained on collected and public datasets, evaluated against external baselines (StyleTTS 2, CosyVoice, VISinger 2, TCSinger) using MOS, CMOS, FFE, and cosine similarity, and ablated component-by-component. No equation in the paper defines a predicted quantity as a function of the fitted quantity used to produce it. The BBC Encoder's duration prediction, Custom Audio Encoder contrastive loss, and flow-matching objective are standard training losses rather than masked re-statements of the evaluation metrics. The paper does cite prior work by the same authors (TCSinger, GTSinger, StyleSinger), but these citations supply datasets, baseline systems, or methodological context; they are not invoked as uniqueness theorems or as the sole justification for the central outperformance claim. The modified baselines in Section 4.1 are underspecified, and hyperparameter choices (gamma=3, four experts) are tuned via CMOS evaluation, but these are experimental-fairness and reporting concerns, not circularity: the comparisons remain against independently implemented systems and the claimed margins are not forced by construction. Accordingly, the paper receives a circularity score of 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper is an empirical system paper. It introduces no new physical entities but relies on several domain assumptions about data quality, representation alignment, and evaluation metrics. The free parameters are hyperparameters chosen by hand or by tuning on evaluation outcomes, rather than quantities derived from first principles.

free parameters (4)
  • Boundary mask length m = 8
    Set manually in Section 3.2 based on compression rate and sample rate; controls how many frames are masked at each phoneme/note boundary and is not swept.
  • CFG scale gamma = 3
    Selected via CMOS comparisons in Appendix E.1; the paper does not state whether a separate validation set was used, so this may be tuned on the test distribution.
  • Number of experts per Cus-MOE group = 4
    Chosen from a sweep in Appendix E.2 balancing quality vs. compute; performance plateaus after 4, indicating tuning on evaluation data.
  • Load-balancing loss weight alpha = 0.1
    Set in Appendix A.4 to control regularization strength; no sensitivity analysis is provided.
assumptions (6)
  • standard math Flow matching and rectified flow interpolation with an Euler solver produce high-quality samples at 25 inference steps.
    Used as the generative backbone (Appendix A.3); the paper relies on prior flow-matching theory without new derivation.
  • domain assumption A single contrastively learned embedding can align singing, speech, and textual prompts while retaining multi-level style and disentangling content.
    Section 3.3 defines three contrast types; the assumption that this yields transferable style representations is only tested indirectly via downstream MOS.
  • domain assumption Masking m frames at phoneme/note boundaries (BBC Encoder) improves robustness to alignment errors without degrading content fidelity.
    Section 3.2 motivates this and Section 4.3 ablates it, but there is no separate measure of content preservation.
  • domain assumption Manual style annotations (gender, vocal range, emotion, vocal technique) are consistent and sufficient to train textual prompt control.
    Introduced in Section 4.1 and acknowledged as a limitation in Section 6; annotation errors propagate through the textual encoder.
  • domain assumption WavLM fine-tuned speaker embeddings and F0 Frame Error are valid objective proxies for singer similarity and synthesis quality.
    Appendix C.2 uses them without reporting variance or correlation with subjective scores.
  • domain assumption The 50 hours of newly collected singing data are clean and correctly annotated.
    Section B states they were collected and annotated, but the data are not released or auditable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TCSinger 2: Customizable Multilingual Zero-shot Singing Voice Synthesis." pith.science (2026). https://pith.science/paper/IXJWZB2F

@misc{pith2026250514910,
  author       = {Pith},
  title        = {Pith review of: TCSinger 2: Customizable Multilingual Zero-shot Singing Voice Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IXJWZB2F}},
  note         = {Machine review of arXiv:2505.14910}
}
read the original abstract

Customizable multilingual zero-shot singing voice synthesis (SVS) has various potential applications in music composition and short video dubbing. However, existing SVS models overly depend on phoneme and note boundary annotations, limiting their robustness in zero-shot scenarios and producing poor transitions between phonemes and notes. Moreover, they also lack effective multi-level style control via diverse prompts. To overcome these challenges, we introduce TCSinger 2, a multi-task multilingual zero-shot SVS model with style transfer and style control based on various prompts. TCSinger 2 mainly includes three key modules: 1) Blurred Boundary Content (BBC) Encoder, predicts duration, extends content embedding, and applies masking to the boundaries to enable smooth transitions. 2) Custom Audio Encoder, uses contrastive learning to extract aligned representations from singing, speech, and textual prompts. 3) Flow-based Custom Transformer, leverages Cus-MOE, with F0 supervision, enhancing both the synthesis quality and style modeling of the generated singing voice. Experimental results show that TCSinger 2 outperforms baseline models in both subjective and objective metrics across multiple related tasks. Singing voice samples are available at https://aaronz345.github.io/TCSinger2Demo/.

Figures

Figures reproduced from arXiv: 2505.14910 by the authors.

Figure 1
Figure 1. The architecture of TCSinger 2. BBC Encoder denotes Blurred Boundary Content Encoder. Figure (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The architecture of Custom Audio Encoder and Cus-MOE. In Figure (a), different encoders extract aligned [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. visualizations of style control. Figure (b) shows more F0 fluctuation than (a), highlighting vibrato. Figure [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The instructions for our subjective evaluation on MOS. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Conan: A Chunkwise Online Network for Zero-Shot Adaptive Voice Conversion

    eess.AS 2025-07 conditional novelty 5.0 of 10

    Conan achieves chunkwise online zero-shot voice conversion, preserving source content while adopting the reference speaker's timbre and style, with a latency as low as 37 milliseconds.

Reference graph

Works this paper leans on

46 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449--12460

    2020

  2. [2]

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. 2022. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505--1518

    2022

  3. [3]

    Yin-Ping Cho, Yu Tsao, Hsin-Min Wang, and Yi-Wen Liu. 2022. Mandarin singing voice synthesis with denoising diffusion probabilistic wasserstein gan. In 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 1956--1963. IEEE

    work page 2022

  4. [4]

    Seungwoo Choi, Seungju Han, Dongyoung Kim, and Sungjoo Ha. 2020. Attentron: Few-shot text-to-speech utilizing attention-based variable-length embedding. arXiv preprint arXiv:2005.08484

    work page Pith review arXiv 2020

  5. [5]

    Soonbeom Choi and Juhan Nam. 2022. A melody-unsupervision model for singing voice synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7242--7246. IEEE

    work page 2022

  6. [6]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1--53

    2024

  7. [7]

    Jianwei Cui, Yu Gu, Chao Weng, Jie Zhang, Liping Chen, and Lirong Dai. 2024. Sifisinger: A high-fidelity end-to-end singing voice synthesizer based on source-filter model. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11126--11130. IEEE

    work page 2024

  8. [8]

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. 2024. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. arXiv preprint arXiv:2407.05407

    arXiv 2024

Show all 46 references
  1. [9]

    William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1--39

    2022

  2. [10]

    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

    2025 arXiv

  3. [11]

    Rongjie Huang, Feiyang Chen, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. 2021. Multi-singer: Fast multi-singer singing voice vocoder with a large-scale corpus. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3945--3954

    2021

  4. [12]

    Ziyue Jiang, Yi Ren, Ruiqi Li, Shengpeng Ji, Boyang Zhang, Zhenhui Ye, Chen Zhang, Bai Jionghao, Xiaoda Yang, Jialong Zuo, et al. 2025. Megatts 3: Sparse alignment enhanced latent diffusion transformer for zero-shot speech synthesis. arXiv preprint arXiv:2502.18924

    2025 arXiv

  5. [13]

    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

  6. [14]

    Tae-Woo Kim, Min-Su Kang, and Gyeong-Hoon Lee. 2022. Adversarial multi-task learning for disentangling timbre and pitch in singing voice synthesis. arXiv preprint arXiv:2206.11558

    2022 arXiv

  7. [15]

    Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  8. [16]

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in neural information processing systems, 33:17022--17033

    2020

  9. [17]

    Neeraj Kumar, Srishti Goel, Ankur Narang, and Brejesh Lall. 2021. Normalization driven zero-shot multi-speaker speech synthesis. In Interspeech, pages 1354--1358

    2021

  10. [18]

    Ruiqi Li, Yu Zhang, Yongqi Wang, Zhiqing Hong, Rongjie Huang, and Zhou Zhao. 2024 a . Robust singing voice transcription serves synthesis. arXiv preprint arXiv:2405.09940

    2024 arXiv

  11. [19]

    Yinghao Aaron Li, Cong Han, and Nima Mesgarani. 2022. Styletts: A style-based generative model for natural and diverse text-to-speech synthesis. arXiv preprint arXiv:2205.15439

    2022 arXiv

  12. [20]

    Yinghao Aaron Li, Cong Han, Vinay Raghavan, Gavin Mischler, and Nima Mesgarani. 2024 b . Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models. Advances in Neural Information Processing Systems, 36

    2024

  13. [21]

    Jinglin Liu, Chengxi Li, Yi Ren, Zhiying Zhu, and Zhou Zhao. 2022 a . Learning the beauty in songs: Neural singing voice beautifier. arXiv preprint arXiv:2202.13277

    2022 arXiv

  14. [22]

    Xingchao Liu, Chengyue Gong, et al. 2022 b . Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations

    2022

  15. [23]

    Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. 2017. Least squares generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2794--2802

    2017

  16. [24]

    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

  17. [25]

    Xiaonan Nie, Xupeng Miao, Shijie Cao, Lingxiao Ma, Qibin Liu, Jilong Xue, Youshan Miao, Yi Liu, Zhi Yang, and Bin Cui. 2021. Evomoe: An evolutional mixture-of-experts training framework via dense-to-sparse gate. arXiv preprint arXiv:2112.14397

    2021 arXiv

  18. [26]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195--4205

    2023

  19. [27]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  20. [28]

    RJ Skerry-Ryan, Eric Battenberg, Ying Xiao, Yuxuan Wang, Daisy Stanton, Joel Shor, Ron Weiss, Rob Clark, and Rif A Saurous. 2018. Towards end-to-end prosody transfer for expressive speech synthesis with tacotron. In international conference on machine learning, pages 4693--4702. PMLR

    2018

  21. [29]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063

    2024

  22. [30]

    Michael Wagner and Duane G Watson. 2010. Experimental and theoretical advances in prosody: A review. Language and cognitive processes, 25(7-9):905--945

    2010

  23. [31]

    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

  24. [32]

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

    2022 arXiv

  25. [33]

    Julian Za di, Hugo Seut \'e , BV Niekerk, and M Carbonneau. 2021. Daft-exprt: Robust prosody transfer across speakers for expressive speech synthesis. arXiv preprint arXiv:2108.02271

    2021 arXiv

  26. [34]

    Biao Zhang and Rico Sennrich. 2019. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32

    2019

  27. [35]

    Lichao Zhang, Ruiqi Li, Shoutong Wang, Liqun Deng, Jinglin Liu, Yi Ren, Jinzheng He, Rongjie Huang, Jieming Zhu, Xiao Chen, et al. 2022 a . M4singer: A multi-style, multi-singer and musical score provided mandarin singing corpus. Advances in Neural Information Processing Syste...

    2022

  28. [36]

    Ruiyuan Zhang, Yuyao Chen, Jiaxiang Liu, Dianbing Xi, Yuchi Huo, Jie Liu, and Chao Wu. 2025 a . Sgw-based multi-task learning in vision tasks. In Asian Conference on Computer Vision, pages 124--141. Springer

    2025

  29. [37]

    Ruiyuan Zhang, Jiaxiang Liu, Zexi Li, Hao Dong, Jie Fu, and Chao Wu. 2024 a . Scalable geometric fracture assembly via co-creation space among assemblers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7269--7277

    2024

  30. [38]

    Ruiyuan Zhang, Qi Wang, Jiaxiang Liu, Yu Zhang, Yuchi Huo, and Chao Wu. 2025 b . Leveraging pretrained diffusion models for zero-shot part assembly. arXiv preprint arXiv:2505.00426

    2025 arXiv

  31. [39]

    Yongmao Zhang, Jian Cong, Heyang Xue, Lei Xie, Pengcheng Zhu, and Mengxiao Bi. 2022 b . 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 Processin...

    2022

  32. [40]

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

    2025

  33. [41]

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

    2025

  34. [42]

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

    2024

  35. [43]

    Yu Zhang, Ziyue Jiang, Ruiqi Li, Changhao Pan, Jinzheng He, Rongjie Huang, Chuxin Wang, and Zhou Zhao. 2024 c . Tcsinger: Zero-shot singing voice synthesis with style transfer and multi-level style control. In Proceedings of the 2024 Conference on Empirical Methods in Natural ...

    2024

  36. [44]

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

    2024

  37. [45]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.