pith. sign in

arxiv: 2606.09295 · v1 · pith:ORNERLVYnew · submitted 2026-06-08 · 💻 cs.CL

N\"ushuVoice: Reviving the Voice of Endangered N\"ushu with Pitch-Aware Text-to-Speech

Hongkun Yang , Xinhui Yi , Xiyan Zhao , Yibo Meng , Lionel Z. Wang , Lixu Wang , Yaqi Zhang , Ruiqi Chen
show 8 more authors
Xuanyue Zhao Lanxin Zhang Yu Zeng Weijia Chu Yiming Ma Chenyu Liu Jianghao Lin Xin Xu
This is my paper

Pith reviewed 2026-06-27 16:38 UTC · model grok-4.3

classification 💻 cs.CL
keywords Nüshutext-to-speechlow-resource TTSpitch conditioningendangered languageVITSprosodyphonetic script
0
0 comments X

The pith

Nüshu-PitchVITS conditions a VITS model on five-level pitch notation to synthesize speech from scarce archival recordings of the endangered script.

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

The paper creates the first TTS benchmark and dataset for Nüshu by aligning Unicode text, phonetic transcriptions, Chinese translations, and limited archival audio. It proposes Nüshu-PitchVITS, which feeds the script's explicit five-level pitch marks into a VITS architecture as an inductive bias for prosody. Experiments demonstrate gains over standard baselines in spectral fidelity, pitch accuracy, and listener intelligibility ratings. The work addresses the gap between mostly syllable-level recordings and the need for sentence-level synthesis in this low-resource setting.

Core claim

Nüshu-PitchVITS, an F0-conditioned VITS framework that leverages Nüshu's five-level pitch notation as an explicit prosodic inductive bias, outperforms strong TTS baselines in spectral fidelity, pitch reconstruction, and human-rated intelligibility.

What carries the argument

Nüshu-PitchVITS: an F0-conditioned VITS model that incorporates the five-level pitch notation extracted from Nüshu as an explicit prosodic bias to guide synthesis when sentence-level data is scarce.

If this is right

  • The pitch-conditioned model produces higher-fidelity audio than unconditioned baselines under the same data constraints.
  • The released sentence-level dataset of aligned Nüshu text and recordings supports further low-resource TTS experiments.
  • Explicit use of the five-level pitch marks compensates for the absence of natural sentence-level utterances in the archives.
  • Human-rated intelligibility improves, indicating the synthesized speech better preserves usable pronunciation.

Where Pith is reading between the lines

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

  • The same explicit pitch-bias approach might transfer to other tonal scripts or languages where archival data consists mainly of isolated syllables.
  • Future work could test whether the five-level notation generalizes to unseen Nüshu sentences without additional alignment effort.
  • If more recordings become available, the model could be fine-tuned to produce longer, more varied prosody while retaining the pitch prior.

Load-bearing premise

The five-level pitch notation extracted from Nüshu can be aligned reliably with the limited archival recordings and functions as an effective explicit prosodic inductive bias inside the VITS architecture.

What would settle it

An ablation experiment or listening test in which removing the pitch conditioning produces no measurable drop in spectral or intelligibility metrics on the Nüshu test set would falsify the utility of the bias.

Figures

Figures reproduced from arXiv: 2606.09295 by Chenyu Liu, Hongkun Yang, Jianghao Lin, Lanxin Zhang, Lionel Z. Wang, Lixu Wang, Ruiqi Chen, Weijia Chu, Xinhui Yi, Xin Xu, Xiyan Zhao, Xuanyue Zhao, Yaqi Zhang, Yibo Meng, Yiming Ma, Yu Zeng.

Figure 1
Figure 1. Figure 1: A simplified, stylized illustration of our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the construction pipeline of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed Nüshu-PitchVITS framework. (a) During training, the text encoder produces phoneme-level hidden states, which are expanded to the frame level by MAS. A pitch predictor estimates the frame-level log-F0 sequence and injects the projected pitch representation into the latent acoustic representation before waveform decoding. The model is optimized with reconstruction, KL-divergence, dur… view at source ↗
Figure 4
Figure 4. Figure 4: Mel-spectrogram comparison of two Nüshu utterances. The left column corresponds to an utter￾ance with IPA transcription [iou13 i5 va33 si44 suow44 mai42 tseng42 ku21 lu21], and the right column cor￾responds to [ciang13 kou21 njyu33 tie42 pwe5 fwe13 ngu13]. 7 Conclusion This paper presents the first dedicated study on text￾to-audio modeling for Nüshu, an endangered pho￾netic script with unique linguistic an… view at source ↗
read the original abstract

N\"ushu is an endangered phonetic script historically used by women in Jiangyong County, southern Hunan, China. While existing computational studies of N\"ushu mainly focus on textual digitization and visual recognition, the acoustic reconstruction of its authentic pronunciation remains largely unexplored. Building a N\"ushu text-to-speech (TTS) system is particularly challenging because available recordings are extremely limited and mostly consist of isolated syllable-level pronunciations rather than natural sentence-level utterances. In this work, we introduce N\"ushuVoice, the first TTS benchmark for N\"ushu. We construct a sentence-level N\"ushu text-to-audio dataset that aligns standardized Unicode N\"ushu text, phonetic transcriptions, standard Chinese translations, and archival recordings. To synthesize speech under this extreme low-resource setting, we propose N\"ushu-PitchVITS, an F0-conditioned VITS framework that leverages N\"ushu's five-level pitch notation as an explicit prosodic inductive bias. Experimental results show that N\"ushu-PitchVITS outperforms strong TTS baselines in spectral fidelity, pitch reconstruction, and human-rated intelligibility. We publicly release the dataset and code at: https://anonymous.4open.science/r/Nvshu-TTS-2EB6.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces NüshuVoice as the first TTS benchmark for the endangered Nüshu script. It constructs a sentence-level dataset aligning standardized Unicode Nüshu text, phonetic transcriptions, Chinese translations, and limited archival recordings (mostly isolated syllables). It proposes Nüshu-PitchVITS, an F0-conditioned VITS model that incorporates Nüshu's five-level pitch notation as an explicit prosodic inductive bias, and claims this outperforms strong TTS baselines in spectral fidelity, pitch reconstruction, and human-rated intelligibility. The dataset and code are released publicly.

Significance. If the results hold, this would be a meaningful contribution to endangered language preservation by enabling acoustic reconstruction of Nüshu pronunciation, where prior work focused only on textual aspects. The explicit use of domain-specific pitch notation as inductive bias in extreme low-resource TTS is a potentially generalizable idea, and the public release of the dataset and code supports reproducibility and follow-on work.

major comments (2)
  1. [Abstract] Abstract and experimental results section: the central claim of outperformance in spectral fidelity, pitch reconstruction, and intelligibility is stated without any numerical metrics, baseline identities, data-split details, or statistical tests, so the strength of evidence cannot be assessed from the provided text.
  2. [Dataset construction] Dataset construction section: the five-level pitch notation alignment with limited archival recordings (mostly isolated syllable pronunciations) is load-bearing for the claim that the F0 conditioning delivers genuine prosodic gains; without explicit details on extraction, alignment procedure, validation against recordings, or error rates, it is unclear whether misalignment or noise in the pitch-to-F0 mapping undermines the inductive bias.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential contribution to endangered language preservation. We address each major comment below and will revise the manuscript to improve clarity and evidence presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract and experimental results section: the central claim of outperformance in spectral fidelity, pitch reconstruction, and intelligibility is stated without any numerical metrics, baseline identities, data-split details, or statistical tests, so the strength of evidence cannot be assessed from the provided text.

    Authors: We agree that the abstract would be strengthened by including key quantitative results. In the revised version we will update the abstract to report specific metrics (e.g., MCD, F0 RMSE, and intelligibility MOS scores), name the baselines, briefly describe the train/test splits, and note any statistical tests performed. The experimental results section already contains these details; the revision will ensure they are summarized concisely in the abstract as well. revision: yes

  2. Referee: [Dataset construction] Dataset construction section: the five-level pitch notation alignment with limited archival recordings (mostly isolated syllable pronunciations) is load-bearing for the claim that the F0 conditioning delivers genuine prosodic gains; without explicit details on extraction, alignment procedure, validation against recordings, or error rates, it is unclear whether misalignment or noise in the pitch-to-F0 mapping undermines the inductive bias.

    Authors: We acknowledge that additional explicit details on the pitch alignment process would help readers evaluate the reliability of the inductive bias. The manuscript describes the use of the five-level notation and its alignment to the archival recordings, but we will expand the dataset construction section to specify the F0 extraction method, the precise alignment procedure between pitch levels and recorded contours, the validation steps taken against the available syllable recordings, and any observed error rates or quality controls. This will clarify that the mapping supports the reported prosodic gains. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper's central claim is an empirical result: Nüshu-PitchVITS outperforms baselines in spectral fidelity, pitch reconstruction, and intelligibility when using Nüshu's historical five-level pitch notation as an explicit conditioning signal inside VITS. This conditioning draws on external domain knowledge of the script rather than any quantity fitted or defined from the target evaluation metrics. No equations, predictions, or uniqueness claims reduce to self-defined inputs, fitted parameters renamed as outputs, or self-citation chains. The dataset construction and model architecture are presented as independent engineering steps whose effectiveness is tested against external baselines.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the representativeness of the constructed sentence-level dataset from syllable recordings and on the assumption that five-level pitch labels transfer usefully as conditioning inside VITS; no free parameters or invented entities are introduced beyond standard model hyperparameters.

axioms (1)
  • domain assumption VITS architecture remains stable and effective when conditioned on discrete pitch levels under extreme data scarcity
    Invoked when the authors add F0 conditioning as an inductive bias without additional validation experiments described in the abstract.

pith-pipeline@v0.9.1-grok · 5811 in / 1229 out tokens · 28763 ms · 2026-06-27T16:38:02.927557+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    Proceedings of IEEE pacific rim conference on communications computers and signal processing , volume=

    Mel-cepstral distance measure for objective speech quality assessment , author=. Proceedings of IEEE pacific rim conference on communications computers and signal processing , volume=. 1993 , organization=

    1993

  9. [9]

    1996 , publisher=

    Methods for subjective determination of transmission quality , author=. 1996 , publisher=

    1996

  10. [10]

    International conference on machine learning , pages=

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  11. [11]

    Introduction: Cultures of authenticity , author=

  12. [12]

    Sun, Yuqian and Tang, Yuying and Gao, Ze and Pan, Zhijun and Xu, Chuyan and Chen, Yurou and Qian, Kejiang and Wang, Zhigang and Braud, Tristan and Lee, Chang Hee and others , booktitle=. AI N

  13. [13]

    arXiv preprint arXiv:2412.00218 , year=

    NushuRescue: Revitalization of the Endangered Nushu Language with AI , author=. arXiv preprint arXiv:2412.00218 , year=

    arXiv

  14. [14]

    Recontextualizing revitalization: A mixed media approach to reviving the n

    Yang, Ivory and Guo, Xiaobo and Wang, Yuxin and Zhang, Hefan and Jia, Yaning and Dinauer, William and Vosoughi, Soroush , booktitle=. Recontextualizing revitalization: A mixed media approach to reviving the n

  15. [15]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  16. [16]

    arXiv preprint arXiv:2006.04558 , year=

    Fastspeech 2: Fast and high-quality end-to-end text to speech , author=. arXiv preprint arXiv:2006.04558 , year=

    arXiv 2006

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Glow-tts: A generative flow for text-to-speech via monotonic alignment search , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    Parameter-Efficient Fine-Tuning for Low-Resource Text-to-Speech via Cross-Lingual Continual Learning , author=. Proc. Interspeech , volume=

  19. [19]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    BnTTS: Few-shot speaker adaptation in low-resource setting , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  20. [20]

    Advancing Women in Leadership Journal , volume=

    The Movement from Secret Acts of Defiance to Manifestation of Women's Empowerment , author=. Advancing Women in Leadership Journal , volume=

  21. [21]

    1995 , publisher=

    Nushu (Chinese women's script) literacy and literature , author=. 1995 , publisher=

    1995

  22. [22]

    Nyushu character unification——theory and rules , author=

  23. [23]

    arXiv preprint arXiv:2004.03136 , year=

    g2pm: A neural grapheme-to-phoneme conversion package for mandarin chinese based on a new open benchmark dataset , author=. arXiv preprint arXiv:2004.03136 , year=

    arXiv 2004

  24. [24]

    2018 IEEE international conference on acoustics, speech and signal processing (ICASSP) , pages=

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions , author=. 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP) , pages=. 2018 , organization=

    2018

  25. [25]

    arXiv preprint arXiv:2106.15561 , year=

    A survey on neural speech synthesis , author=. arXiv preprint arXiv:2106.15561 , year=

    arXiv

  26. [26]

    , author=

    A Systematic Review and Analysis of Multilingual Data Strategies in Text-to-Speech for Low-Resource Languages. , author=. Interspeech , pages=

  27. [27]

    Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Lrspeech: Extremely low-resource speech synthesis and recognition , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

  28. [28]

    ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Unsupervised Pre-Training for Data-Efficient Text-to-Speech on Low Resource Languages , author=. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2023 , organization=

    2023

  29. [29]

    2023 , eprint=

    Unsupervised Pre-Training For Data-Efficient Text-to-Speech On Low Resource Languages , author=. 2023 , eprint=

    2023

  30. [30]

    The Oxford handbook of Chinese linguistics , pages=

    Wu dialect , author=. The Oxford handbook of Chinese linguistics , pages=. 2015 , publisher=

    2015

  31. [31]

    2024 , school=

    Jiangyong Nvshu in China: the Gender, Signifier, Signified, and Sustainable Inheritance Path of the Local Culture , author=. 2024 , school=

    2024

  32. [32]

    History, Characteristics, and Modern Vitality of N

    Congrong, Li , journal=. History, Characteristics, and Modern Vitality of N. 2024 , publisher=

    2024

  33. [33]

    arXiv preprint arXiv:2402.08093 , year=

    Base tts: Lessons from building a billion-parameter text-to-speech model on 100k hours of data , author=. arXiv preprint arXiv:2402.08093 , year=

    arXiv

  34. [34]

    International Conference on Multi-modal Information Analytics , pages=

    Intangible Cultural Heritage Protection Based on MAR Algorithm , author=. International Conference on Multi-modal Information Analytics , pages=. 2022 , organization=

    2022

  35. [35]

    Proceedings of the Eight Workshop on the Use of Computational Methods in the Study of Endangered Languages , pages=

    Evaluating Indigenous language speech synthesis for education: A participatory design workshop on Ojibwe text-to-speech , author=. Proceedings of the Eight Workshop on the Use of Computational Methods in the Study of Endangered Languages , pages=

  36. [36]

    The world’s first South S

    Hiovain-Asikainen, Katri and Kj. The world’s first South S. Proceedings of the 10th International Workshop on Computational Linguistics for Uralic Languages , pages=

  37. [37]

    Falcini, Giulia , journal=. N