Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Technical report: Impact of Duration Prediction on Speaker-specific TTS for Indian Languages

T0 review · 3 major / 6 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 The choice of duration predictor in a zero-shot Indian-language TTS system trades intelligibility against speaker similarity, and which side wins depends on the language.

desk verdict Useful comparison of duration predictors for Indian-language TTS, but possible train/test overlap and missing statistical rigor make the central claims provisional. read the letter →

arxiv 2507.16875 v1 pith:757ONTIM submitted 2025-07-22 eess.AS cs.LG

Isha Pandey , Pranav Gaikwad , Amruta Parulekar , Ganesh Ramakrishnan This is my paper
classification eess.AScs.LG
keywords durationpredictionzero-shotTTSIndianlanguagescontinuousnormalizingflowspeechinfillingspeakersimilarityintelligibilitylow-resourcesynthesis
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

This technical report asks whether the way a TTS system predicts phoneme durations changes the quality of zero-shot, speaker-specific speech in low-resource Indian languages. Training a Voicebox-style continuous-normalizing-flow model on roughly 1,000 to 3,000 hours per language, the authors compare an infilling-style duration predictor (conditioned on text and context durations) with a speaker-prompted predictor (conditioned on a three-second speech prompt via cross-attention). They find language-dependent trade-offs: the speaker-prompted predictor improves Word Error Rate for Tamil and Bengali, but degrades it sharply for Marathi, while speaker similarity generally improves with the speaker-prompted predictor. If correct, the result means duration prediction cannot be treated as a fixed module; its design should be selected per language and per task, and intelligibility and speaker fidelity may not be jointly optimized by one strategy.

What carries the argument

The load-bearing mechanism is the pairing of two duration predictors with one fixed CNF audio model. The infilling predictor models $q(l \mid y, l_{\text{ctx}})$ by regressing masked log-durations from text and context duration sequences using a Transformer (51M parameters). The speaker-prompted predictor models $q(l \mid y, x_{\text{ctx}})$ instead: a speech-prompted text encoder (84M parameters) projects a three-second mel-spectrogram prompt and text into a shared space, applies cross-attention from text tokens to speech frames, and feeds the speaker-conditioned representation to a convolutional duration predictor. Both feed the same 103M-parameter Transformer CNF audio model trained with masked flow matching (0.9 weight on masked frames), so any difference in output is attributable to the duration strategy.

What would settle it

Generate the same Marathi test sentences with both duration predictors, transcribe with two independent ASR systems, and compute WER with confidence intervals; if the Marathi gap between infilling and speaker-prompted durations (0.224 versus 0.365 overall) disappears or reverses under the second ASR, the claimed language-dependent degradation is a metric artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that the duration-prediction strategy is a decisive, language-dependent design choice for zero-shot, speaker-specific TTS in low-resource Indian languages. Using the same Voicebox-style CNF audio model, the authors compare an infilling-style duration predictor that conditions on text and context durations (requiring forced alignment) with a speaker-prompted predictor that conditions on a three-second speech prompt via cross-attention, trained with weak alignment supervision. Across Tamil, Telugu, Bengali, Hindi, and Marathi, the speaker-prompted predictor lowers Word Error Rate relative to infilling for Tamil and Bengali (e.g., Tamil overall WER 0.312 to 0.295) but raises it sharply for Marathi (0.224 to 0.365), while Sim-o speaker similarity improves for Tamil, Telugu, and Bengali. The paper concludes with a general trend: speaker-prompted durations tend to preserve speaker characteristics better, while infilling durations tend to preserve intelligibility better, making the choice a task- and language-dependent trade-off.

Load-bearing premise

The paper relies on WER from IndicConformer and cosine similarity from an ECAPA-TDNN speaker encoder trained on VoxCeleb2 as accurate cross-lingual measures for five Indian languages, without reporting confidence intervals or validating these metrics on the target languages; if these metrics are biased or noisy, the observed language-dependent trade-offs could be artifacts.

Editorial extensions

If this is right

  • In Tamil, the speaker-prompted predictor improves both WER (overall 0.312 to 0.295) and Sim-o (0.683 to 0.693) over infilling, so it is the safer choice for that language.
  • For Marathi, switching from infilling to speaker-prompted durations raises overall WER from 0.224 to 0.365, so a language-by-language selection is needed rather than a universal default.
  • Human QMOS and SMOS ratings for Hindi and Tamil confirm that the speaker-prompted predictor matches or beats infilling on naturalness and similarity, supporting that objective-metric trends are not purely ASR artifacts.
  • The choice of duration predictor cannot be decoupled from the language's prosodic structure; future systems should expose duration strategy as a tunable component.

Reading between the lines

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

  • The paper's general trend implies a trade-off frontier: speaker-prompted durations inject prosodic and speaker cues at the cost of segmental alignment accuracy, while infilling durations preserve timing at the cost of speaker identity. A hybrid predictor that interpolates between prompt-derived and text-context-derived duration estimates could let a system choose its position on this frontier per
  • The Marathi-specific WER collapse suggests the speaker-prompted predictor may be sensitive to how well the prompt's prosody matches the target sentence's rhythm; languages with higher intra-speaker durational variance would amplify any mismatch. Breaking down Marathi errors by dialect, speaking rate, and prompt-text rhythm distance would test this hypothesis.
  • Because both objective metrics come from models not validated on these target languages (IndicConformer for ASR, ECAPA-TDNN trained on VoxCeleb2 for speaker similarity), the language-dependent ranking could partly reflect metric bias. Re-running the comparison with a second ASR and a language-specific speaker-verification model would show whether the trade-off pattern survives.
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

3 major / 6 minor

Summary. This paper compares two duration prediction strategies in a Voicebox-inspired continuous normalizing flow (CNF) speech infilling model trained for five Indian languages (Hindi, Marathi, Tamil, Telugu, Bengali). The first strategy follows Voicebox in predicting durations from text and context durations; the second follows PFlow in predicting durations from a three-second speaker prompt via cross-attention. The authors evaluate intelligibility with WER (IndicConformer), speaker similarity with Sim-o (ECAPA-TDNN), and human SMOS/QMOS on 40 samples per language, using the Vistaar test set. The reported results are mixed: the speaker-prompted predictor improves WER for Tamil and Bengali but is much worse for Marathi, and it improves Sim-o for Tamil, Telugu, and Bengali. The paper concludes that duration prediction strategy is language- and metric-dependent.

Significance. The question is practically important: duration prediction is a component that is easy to swap, and evidence on how it behaves in low-resource Indian languages is scarce. The paper's strengths are the same-audio-model comparison of the two predictors across multiple languages and datasets, the use of both objective and human metrics, and the transparent reporting of data filtering and training budgets. If the central claim is supported, the results would provide useful design guidance. However, the current evidence is not yet conclusive: the lack of train/test deduplication checks, the absence of confidence intervals or significance tests, and the small human-evaluation sample are substantial limitations.

major comments (3)
  1. [§4.1/§4.2 and Appendix C/Table 8] The paper does not report any train/test deduplication. The training corpus (Section 4.1 and Table 1) includes IndicSuperb, MUCS, and FLEURS, and the evaluation set (Appendix C) draws samples from exactly these corpora. Without a stated utterance-level disjointness check (e.g., file-hash or transcript-overlap), the language-dependent gaps in Table 3—for example Marathi IndicSuperb WER of 0.39 for PFlow versus 0.1862 for Infill—cannot be interpreted as causal effects of the duration predictor. The authors should document a deduplication procedure or re-run the comparison on a held-out split that is explicitly disjoint from training.
  2. [§4.5 and Tables 3–5] All objective results are point estimates with no confidence intervals, bootstrap intervals, or significance tests. Several reported differences are small (e.g., Tamil overall WER of 0.31160 for GT, 0.31185 for Infill, and 0.29455 for PFlow; Tamil Sim-o of 0.6925 for PFlow versus 0.6833 for Infill), and the human evaluation uses only 40 samples per language with 10 ratings each. Without measures of uncertainty and inter-annotator agreement, the central trade-off claim is not statistically supported.
  3. [§5.1 and §6] The conclusion that 'the infilling duration predictor led to better intelligibility' as a general trend is not consistent with Table 3, where PFlow improves overall WER for Tamil and Bengali while Infill is better for Hindi, Telugu, and Marathi; the dataset-level results are also mixed (e.g., Marathi FLEURS: PFlow 0.3526 vs Infill 0.348). The authors should either soften the general-trend statement to an observed language-specific pattern or provide a statistical model that accounts for language and dataset effects.
minor comments (6)
  1. [Abstract] The sentence 'While some recent generative approaches choose to omit explicit duration modeling, often at the cost of longer training times.' is a fragment; it should be joined to the following sentence.
  2. [Table 5 caption] The caption states 'Hindi, Tamil and Telugu outputs' but the table shows Hindi, Tamil, and Bengali; the caption should be corrected.
  3. [§4.5.1 and §4.5.2] Both subsections refer to Appendix A for the SMOS and QMOS guidelines, but Appendix A is QMOS and Appendix B is SMOS; the cross-references should be fixed.
  4. [§3.4.2] The placeholder 'PFlow [citation]' should be replaced with the full citation for Kim et al. (2023) at the point of use.
  5. [Various] Minor typos include 'Fluers' in §4.2.1 and Table 1, 'inlcuding' in §2, and 'commpletely' in §4.5.2.
  6. [Appendix C] The text says 'We use the Vistaar test set' but Table 8 lists datasets (IndicSuperb, MUCS, FLEURS) rather than the Vistaar subset name; clarify how the Vistaar test set is constructed and how the per-dataset counts follow from the described filtering.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons of trained duration-prediction strategies, not derivations from fitted parameters or self-citation chains.

full rationale

The paper's claimed derivation chain is empirical rather than definitional. It adapts a CNF-based audio model (Section 3.3) and compares two duration predictors: a Voicebox-style infilling predictor (Section 3.4.1) and a PFlow-inspired speaker-prompted predictor (Section 3.4.2). Both are trained as separate modules and evaluated on held-out Vistaar test sets using WER, Sim-o, and human scores. No equation defines one system's output in terms of the other, and no fitted parameter is renamed as a prediction: the duration predictors are trained with MSE/CFM losses against forced-alignment targets and then compared by downstream metrics. The central result (a language-dependent trade-off between intelligibility and speaker similarity) is an observed pattern in Tables 3-5, not a consequence of the definitions of the two predictors. The references to Voicebox, PFlow, and other prior work are architectural inspirations or standard background, not load-bearing self-citations or imported uniqueness theorems. The skeptic's concerns about possible train/test overlap because the Vistaar test set is drawn from corpora also used for training, and about ASR and speaker-embedding metric reliability, are experimental validity and leakage risks rather than circularity in the defined sense; the paper text does not exhibit an equation-level or definition-level reduction of any claimed result to its own inputs. Therefore no circular step is identified.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central comparison rests mainly on the reliability of its evaluation metrics and the representativeness of its filtered data. These are domain assumptions not independently validated in the paper; the loss weighting and prompt length are hand-chosen hyperparameters that could affect both duration predictors differently.

free parameters (5)
  • Masked loss weight = 0.9 masked, 0.1 context
    Weighted loss in the audio model prioritizes masked-frame reconstruction; chosen by hand and used for all languages.
  • Audio filtering WER threshold = 0.2
    This threshold in Section 4.2.2 discards audios with high WER from IndicWhisper, affecting training data composition.
  • Speaker prompt length = 3 seconds
    The speaker-prompted duration predictor uses a 3-second mel segment from context (Section 3.4.2); prompt length is a design choice.
  • Masking schedule = 50% random mask r~U[30,100]; 90% full mask within that branch; 10% no mask
    Section 4.4 describes a probabilistic masking strategy that controls the training distribution of context and missing audio.
  • CTC add-back threshold = 0.9
    Discarded audios with average CTC alignment score above 0.9 are added back to the training set (Section 4.2.2).
assumptions (5)
  • standard math Optimal transport conditional flow matching is a valid training objective for CNFs (Lipman et al. 2023).
    The audio model in Section 3.3 relies on this framework without re-deriving it.
  • domain assumption IndicConformer and IndicWhisper yield accurate transcriptions across all five Indian languages.
    Used for data filtering (Section 4.2.2) and WER evaluation (Section 4.5.2); no per-language accuracy analysis is provided.
  • domain assumption ECAPA-TDNN trained on VoxCeleb2 produces speaker embeddings that represent Indian voices well enough for Sim-o.
    Section 4.5.1 uses Sim-o to measure speaker similarity; the model was trained on English data.
  • domain assumption Forced alignment provides reliable ground-truth durations in Indian languages.
    The paper uses forced alignment to obtain GT durations while also stating in Section 1 that forced alignments may be unreliable in low-resource contexts.
  • domain assumption The filtered public datasets are representative of the target languages for zero-shot speaker-specific TTS.
    Section 4.1 and 4.2 assume the assembled multi-source corpus covers the linguistic diversity needed; no speaker diversity analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Technical report: Impact of Duration Prediction on Speaker-specific TTS for Indian Languages." pith.science (2026). https://pith.science/paper/757ONTIM

@misc{pith2026250716875,
  author       = {Pith},
  title        = {Pith review of: Technical report: Impact of Duration Prediction on Speaker-specific TTS for Indian Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/757ONTIM}},
  note         = {Machine review of arXiv:2507.16875}
}
read the original abstract

High-quality speech generation for low-resource languages, such as many Indian languages, remains a significant challenge due to limited data and diverse linguistic structures. Duration prediction is a critical component in many speech generation pipelines, playing a key role in modeling prosody and speech rhythm. While some recent generative approaches choose to omit explicit duration modeling, often at the cost of longer training times. We retain and explore this module to better understand its impact in the linguistically rich and data-scarce landscape of India. We train a non-autoregressive Continuous Normalizing Flow (CNF) based speech model using publicly available Indian language data and evaluate multiple duration prediction strategies for zero-shot, speaker-specific generation. Our comparative analysis on speech-infilling tasks reveals nuanced trade-offs: infilling based predictors improve intelligibility in some languages, while speaker-prompted predictors better preserve speaker characteristics in others. These findings inform the design and selection of duration strategies tailored to specific languages and tasks, underscoring the continued value of interpretable components like duration prediction in adapting advanced generative architectures to low-resource, multilingual settings.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 12 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Abhayjeet and 1 others. 2025. 'SYSPIN-S1. 0 Corpus - A TTS Corpus of 900+ hours in nine Indian Languages '

    work page 2025

  4. [4]

    AI4Bharat. 2024. Indicconformer: A suite of asr models for indian languages. https://github.com/AI4Bharat/IndicConformerASR. Accessed: 2025-05-28

    work page 2024

  5. [5]

    Tyers, and Gregor Weber

    Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. 2020. https://arxiv.org/abs/1912.06670 Common voice: A massively-multilingual speech corpus . Preprint, arXiv:1912.06670

    arXiv 2020

  6. [6]

    Kaushal Santosh Bhogale, Abhigyan Raman, Tahir Javed, Sumanth Doddapaneni, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2023 a . Effectiveness of mining audio and text pairs from public data for improving ASR systems for low-resource languages. In ICASSP , pages 1--5. IEEE

    work page 2023

  7. [7]

    Khapra, and Pratyush Kumar

    Kaushal Santosh Bhogale, Sai Sundaresan, Abhigyan Raman, Tahir Javed, Mitesh M. Khapra, and Pratyush Kumar. 2023 b . https://arxiv.org/abs/2305.15386 Vistaar: Diverse benchmarks and training sets for indian language asr . Preprint, arXiv:2305.15386

    arXiv 2023

  8. [8]

    Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. 2018. https://doi.org/10.21437/interspeech.2018-1929 Voxceleb2: Deep speaker recognition . In Interspeech 2018. ISCA

    doi:10.21437/interspeech.2018-1929 2018

Show all 40 references
  1. [9]

    Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. 2022. https://arxiv.org/abs/2205.12446 Fleurs: Few-shot learning evaluation of universal representations of speech . Preprint, arXiv:2205.12446

    2022 arXiv

  2. [10]

    Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. 2020. https://doi.org/10.21437/interspeech.2020-2650 Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification . In Interspeech 2020. ISCA

    2020 doi

  3. [11]

    Anuj Diwan, Rakesh Vaideeswaran, Sanket Shah, Ankita Singh, Srinivasa Raghavan, Shreya Khare, Vinit Unni, Saurabh Vyas, Akash Rajpuria, Chiranjeevi Yarra, Ashish Mittal, Prasanta Kumar Ghosh, Preethi Jyothi, Kalika Bali, Vivek Seshadri, Sunayana Sitaram, Samarth Bharadwaj, Jai...

    2021 doi

  4. [12]

    Tahir Javed, Kaushal Santosh Bhogale, Abhigyan Raman, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. https://arxiv.org/abs/2208.11761 Indicsuperb: A speech processing universal performance benchmark for indian languages . Preprint, arXiv:2208.11761

    2022 arXiv

  5. [13]

    Tahir Javed, Janki Atul Nawale, Eldho Ittan George, Sakshi Joshi, Kaushal Santosh Bhogale, Deovrat Mehendale, Ishvinder Virender Sethi, Aparna Ananthanarayanan, Hafsah Faquih, Pratiti Palit, Sneha Ravishankar, Saranya Sukumaran, Tripura Panchagnula, Sunjay Murali, Kunal Sharad...

    2024 arXiv

  6. [14]

    Jaehyeon Kim, Sungwon Kim, Jungil Kong, and Sungroh Yoon. 2020. https://arxiv.org/abs/2005.11129 Glow-tts: A generative flow for text-to-speech via monotonic alignment search . Preprint, arXiv:2005.11129

    2020 arXiv

  7. [15]

    Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. https://arxiv.org/abs/2106.06103 Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech . Preprint, arXiv:2106.06103

    2021 arXiv

  8. [16]

    Shih, Rohan Badlani, Joao Felipe Santos, Evelina Bakhturina, Mikyas T

    Sungwon Kim, Kevin J. Shih, Rohan Badlani, Joao Felipe Santos, Evelina Bakhturina, Mikyas T. Desta, Rafael Valle, Sungroh Yoon, and Bryan Catanzaro. 2023. https://openreview.net/forum?id=zNA7u7wtIN P-flow: A fast and data-efficient zero-shot TTS through speech prompting . In T...

    2023

  9. [17]

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. https://arxiv.org/abs/2010.05646 Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis . Preprint, arXiv:2010.05646

    2020 arXiv

  10. [18]

    Khapra, and Karthik Nandakumar

    Gokul Karthik Kumar, Praveen S V, Pratyush Kumar, Mitesh M. Khapra, and Karthik Nandakumar. 2023. https://arxiv.org/abs/2211.09536 Towards building text-to-speech systems for the next billion users . Preprint, arXiv:2211.09536

    2023 arXiv

  11. [19]

    Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, and Wei-Ning Hsu. 2023. https://arxiv.org/abs/2306.15687 Voicebox: Text-guided multilingual universal speech generation at scale . Preprint, a...

    2023 arXiv

  12. [20]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2023. https://arxiv.org/abs/2210.02747 Flow matching for generative modeling . Preprint, arXiv:2210.02747

    2023 arXiv

  13. [21]

    Emile Mathieu and Maximilian Nickel. 2020. Riemannian continuous normalizing flows. Advances in neural information processing systems, 33:2503--2515

    2020

  14. [22]

    Chenfeng Miao, Shuang Liang, Zhencheng Liu, Minchuan Chen, Jun Ma, Shaojun Wang, and Jing Xiao. 2020. https://arxiv.org/abs/2012.03500 Efficienttts: An efficient and high-quality text-to-speech architecture . Preprint, arXiv:2012.03500

    2020 arXiv

  15. [23]

    Soumya Priyadarsini Panda, Ajit Kumar Nayak, and Satyananda Champati Rai. 2020. https://api.semanticscholar.org/CorpusID:218978026 A survey on speech synthesis techniques in indian languages . Multimedia Systems, 26:453 -- 478

    2020

  16. [24]

    Nithya R, Malavika S, Jordan F, Arjun Gangwar, Metilda N J, S Umesh, Rithik Sarab, Akhilesh Kumar Dubey, Govind Divakaran, Samudra Vijaya K, and Suryakanth V Gangashetty. 2023. https://arxiv.org/abs/2310.14654 Spring-inx: A multilingual indian language speech corpus by spring ...

    2023 arXiv

  17. [25]

    Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2022. https://arxiv.org/abs/2006.04558 Fastspeech 2: Fast and high-quality end-to-end text to speech . Preprint, arXiv:2006.04558

    2022 arXiv

  18. [26]

    Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2019. https://arxiv.org/abs/1905.09263 Fastspeech: Fast, robust and controllable text to speech . Preprint, arXiv:1905.09263

    2019 arXiv

  19. [27]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. https://arxiv.org/abs/1505.04597 U-net: Convolutional networks for biomedical image segmentation . Preprint, arXiv:1505.04597

    2015 arXiv

  20. [28]

    Ashwin Sankar, Srija Anand, Praveen Srinivasa Varadhan, Sherry Thomas, Mehak Singal, Shridhar Kumar, Deovrat Mehendale, Aditi Krishana, Giri Raju, and Mitesh Khapra. 2024. https://arxiv.org/abs/2409.05356 Indicvoices-r: Unlocking a massive multilingual multi-speaker speech cor...

    2024 arXiv

  21. [29]

    Sanket Shah, Kavya Ranjan Saxena, Kancharana Manideep Bharadwaj, Sharath Adavanne, and Nagaraj Adiga. 2025. IndicST : Indian multilingual translation corpus for evaluating speech large language models. In Proc. ICASSP

    2025

  22. [30]

    Jonathan Shen, Ye Jia, Mike Chrzanowski, Yu Zhang, Isaac Elias, Heiga Zen, and Yonghui Wu. 2021. https://arxiv.org/abs/2010.04301 Non-attentive tacotron: Robust and controllable neural tts synthesis including unsupervised duration modeling . Preprint, arXiv:2010.04301

    2021 arXiv

  23. [31]

    Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A

    Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. 2018. https://arxiv.org/abs/1712.05884 Natural tts synthesis by conditioning w...

    2018 arXiv

  24. [32]

    Shih, Rafael Valle, Rohan Badlani, Adrian Lancucki, Wei Ping, and Bryan Catanzaro

    Kevin J. Shih, Rafael Valle, Rohan Badlani, Adrian Lancucki, Wei Ping, and Bryan Catanzaro. 2021. https://openreview.net/forum?id=0NQwnnwAORi RAD - TTS : Parallel flow-based TTS with robust alignment learning and diverse synthesis . In ICML Workshop on Invertible Neural Networ...

    2021

  25. [33]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2023. https://arxiv.org/abs/2104.09864 Roformer: Enhanced transformer with rotary position embedding . Preprint, arXiv:2104.09864

    2023 arXiv

  26. [34]

    Xu Tan, Jiawei Chen, Haohe Liu, Jian Cong, Chen Zhang, Yanqing Liu, Xi Wang, Yichong Leng, Yuanhao Yi, Lei He, Frank Soong, Tao Qin, Sheng Zhao, and Tie-Yan Liu. 2022. https://arxiv.org/abs/2205.04421 Naturalspeech: End-to-end text to speech synthesis with human-level quality ...

    2022 arXiv

  27. [35]

    VAANI Team. 2025. Vaani: Capturing the language landscape for an inclusive digital india (phase 1). https://vaani.iisc.ac.in/

    2025

  28. [36]

    Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. 2016. https://arxiv.org/abs/1609.03499 Wavenet: A generative model for raw audio . Preprint, arXiv:1609.03499

    2016 arXiv

  29. [37]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. https://arxiv.org/abs/1706.03762 Attention is all you need . Preprint, arXiv:1706.03762

    2023 arXiv

  30. [38]

    Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and Furu Wei. 2023. https://arxiv.org/abs/2301.02111 Neural codec language models are zero-shot text to speech synthesizers . Preprint, ...

    2023 arXiv

  31. [39]

    Yuxuan Wang, Daisy Stanton, Yu Zhang, RJ Skerry-Ryan, Eric Battenberg, Joel Shor, Ying Xiao, Fei Ren, Ye Jia, and Rif A. Saurous. 2018. https://arxiv.org/abs/1803.09017 Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis . Preprint, a...

    2018 arXiv

  32. [40]

    Biao Zhang and Rico Sennrich. 2019. https://arxiv.org/abs/1910.07467 Root mean square layer normalization . Preprint, arXiv:1910.07467

    2019 arXiv

Pith tools

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