Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Adaptive Duration Model for Text Speech Alignment

T0 review · 3 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 DurFormer predicts phoneme durations conditioned on speaking speed and speech scene, delivering 15.05% lower MSE than baselines and more stable zero-shot TTS output.

desk verdict DurFormer is a sensible duration-model integration with consistent reported gains, but the inference-time source of its speed/scene conditioning is never stated, which could make the headline improvements a product of label leakage rather than modeling. read the letter →

arxiv 2507.22612 v2 pith:KAKXCTOP submitted 2025-07-30 cs.SD cs.AIeess.AS

Junjie Cao This is my paper
classification cs.SDcs.AIeess.AS
keywords textspeechalignmentdurationpredictionzero-shottext-to-speechphonemespeakingspeedconditioningsceneflowmatchingsynthesis
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 paper argues that text–speech alignment in non-autoregressive TTS can be improved by treating phoneme duration prediction as a conditional regression with explicit conditioning on speaking speed and speech scene. The proposed model, DurFormer, outputs a mean and variance for the duration sequence rather than a single deterministic value, and fuses semantic text information through cross-attention. On phoneme-level mean squared error it reports a 15.05% improvement at small model size and 11.26% at large size over FastSpeech2, and a roughly 5% word-error-rate reduction over Ratio-Scale when plugged into F5-TTS. The paper also claims the model is more robust when prompt audio and input text differ in speaking circumstance.

What carries the argument

The central object is DurFormer's conditional duration predictor. It consists of an attribute encoder that embeds a five-level speaking-speed factor and a speech-scene factor; a semantic fusion module that uses a pretrained language model's hidden states through an MLP and cross-attention; and a probability module that predicts the mean and variance of a multi-variate Gaussian duration sequence. The training objective is the Gaussian negative log-likelihood in Eq. (6). This machinery turns duration prediction into a conditional distributional regression, so the model can express uncertainty and adapt to speaking conditions.

What would settle it

Re-run the comparison with DurFormer's speed level and speech scene computed only from the prompt or reference audio, never from the target audio; if the 15.05% and 11.26% MSE advantages and the heterogeneous WER gain disappear, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that explicit, quantized conditioning factors—subjective speaking speed (five levels) and objective speech scene—plus semantic text features let a neural duration model predict phoneme-level durations that are closer to an MFA-aligned ground truth and more stable than fixed ratio scaling. DurFormer formulates alignment as minimization of a Gaussian negative log-likelihood over duration sequences, predicting mean and variance, and it is evaluated as a drop-in duration predictor in F5-TTS. The paper reports superior MSE, lower error variance, and lower WER, especially on heterogeneous prompt-target pairs.

Load-bearing premise

The load-bearing premise is that the speaking-speed level and speech-scene attribute are available at inference without peeking at the target duration; the paper never specifies how they are measured, so if they are inferred from the ground-truth durations, the reported advantage is an artifact.

Editorial extensions

If this is right

  • DurFormer can replace Ratio-Scale in F5-TTS to lower WER by roughly 5% and reduce duration error variance.
  • The probability module's mean and variance output allows stochastic duration sampling at inference, which may improve prosodic diversity in downstream TTS.
  • Conditioning on speaking speed level helps both fast and slow speech: removing speed control increases error from 32.13 to 42.21 on a fast set and from 62.23 to 94.29 on a slow set.
  • The model delivers its gains at both small (~150M) and large (~330M) model scales, suggesting the conditioning approach is scale-independent.
  • The reported heterogeneous-sample results indicate the model is more robust than Ratio-Scale when prompt audio and target text come from different speaking circumstances.

Reading between the lines

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

  • If speed level and speech scene are derived reliably from the prompt or reference audio at inference, DurFormer-style conditioning could be added to other non-autoregressive TTS systems as a plug-in front-end without changing the decoder.
  • The variance output could be used to sample multiple duration sequences for the same text, providing a testable way to generate prosodically varied speech from a single model.
  • The phoneme-level MSE gains may translate into larger end-to-end gains on long or out-of-domain text, where fixed ratio scaling is known to drift, but this extension is not explicitly evaluated in the paper.
  • A direct ablation that varies how speed level and speech scene are computed (from prompt audio versus from target durations) would pin down whether the reported robustness comes from the conditioning mechanism itself or from information leakage in the evaluation pipeline.
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 / 5 minor

Summary. The paper proposes DurFormer, a conditional phoneme-duration predictor for TTS alignment. DurFormer conditions on a five-level speaking-speed attribute and a speech-scene attribute through an attribute encoder, fuses text semantics from a pretrained RoBERTa model via cross-attention, and outputs Gaussian mean/variance parameters for the duration sequence (Eqs. (3)-(6)). Experiments on WenetSpeech4TTS and Seed-TTS report phoneme-level MSE improvements of 15.05% (small) and 11.26% (large) over FastSpeech2, WER reductions in F5-TTS (Table 1), and larger gains on 200 heterogeneous prompt-target samples (Table 2).

Significance. If the reported protocol is sound, DurFormer offers a practical way to inject speaking-rate and scene conditions into duration prediction for non-autoregressive TTS, and its distributional (mean+variance) output is a plausible mechanism for improving stability. The paper's attempt to construct heterogeneous prompt-target samples is a useful probe of zero-shot robustness. However, the contribution's central claim currently hinges on the inference-time availability of the two conditioning attributes, which the manuscript never specifies; as written, the claim is not falsifiable from the reported experiments. The paper also does not provide code, data splits, or significance statistics, so the numerical evidence is weaker than the headline percentages suggest.

major comments (3)
  1. [§2.2.1 and §3.1] Section 2.2.1 defines τ_in as five speed levels 'determined by the average phoneme speaking frequency' and τ_ex as a speech scene, and Eq. (3) makes them inputs to the predictor. The manuscript never states how these attributes are obtained when DurFormer is used to predict durations for a target utterance whose speech does not yet exist. If, during the F5-TTS evaluation in Section 3.1, τ_in/τ_ex are computed from the ground-truth alignment of the target audio, then DurFormer is given duration-derived side information at test time while Ratio-Scale, FlowMatching, and FastSpeech2 are not. The reported gains (15.05% MSE at small size, 11.26% at large size; WER 8.33 vs 8.72 and 2.15 vs 2.31 in Table 1; WER 6.57 vs 10.21 in Table 2) would then reflect leakage rather than adaptation. The authors need to specify the exact inference-time source of τ_in/τ_ex (e.g., reference prompt audio, a coarse first-pass duration estimate, or user-provided labels) and re-run the experiments under that protocol; the current text makes the central claim unfalsifiable.
  2. [§3.1 (Table 2)] The construction of the 200 'heterogeneous samples' is not described. The paper says 'speech circumstance differs a lot' but gives no selection criteria, no annotation protocol, and no examples; without this, Table 2's large WER reduction (10.21 to 6.57) cannot be interpreted or replicated. In addition, none of the tables report error bars, confidence intervals, or significance tests, so the smaller differences in Table 1 (e.g., WER 2.15 vs 2.19, MAE 55.43 vs 62.61) may be within run-to-run variability. Please report the number of evaluation utterances per condition, repeated-seed variance, and appropriate significance tests, and describe the heterogeneous-set construction in detail.
  3. [§3.2 (Table 3)] The metric and the numbers in Table 3 appear internally inconsistent. The text defines err = ||D_gt - D_pred||^2 / L, which is a squared error, and states that the duration unit is 10 ms, but Table 3's MSE-Avg values have the same magnitude as the MAE values in Table 1: FastSpeech2-L appears as 62.61 in both tables. Either the tables use different evaluation sets, or the metric definition or unit is wrong. Since the headline '15.05%' and '11.26%' improvements are computed from these values, the authors must clarify the metric, the unit, and whether the same evaluation set is used, and should report the full error distribution rather than only averages.
minor comments (5)
  1. [Abstract and Table 3] There are typographical errors: 'immprovement' in the abstract and '50..94' in Table 3's DurFormer-S MSE-Min entry; please proofread the text and tables.
  2. [Table 4] Table 4 is formatted as one run-on line ('w/o A-E 59.24 w/o S-C 42.21 94.29w/o S-E 58.36'), making it impossible to tell which cells correspond to which ablation and condition; please split the table into proper columns with clear row and column labels.
  3. [Section 4] The conclusion states that DurFormer outperforms baselines 'with respect to efficiency,' but Table 1 shows RTF 0.23 for DurFormer-S versus 0.15 for Ratio-Scale and 0.21 for FastSpeech2-S; the efficiency claim should be qualified to accuracy or removed.
  4. [§2.2.2] The semantic extractor is referred to as 'Pretrained Roberta,' but no citation, model size, or checkpoint is given; please add the exact model used.
  5. [§2.1] The notation for the alignment decomposition, S(Φ)=D N Φ and D N Φ=(...), is under-specified and the equation formatting is garbled; rewriting with standard set/sequence notation would improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Speed-level conditioning is defined from the target duration sequence; ground-truth-derived fast/slow sets in the ablation suggest evaluation uses target-derived τin, making part of the reported gains by construction.

  1. self definitional [Section 2.2.1, Eq. (3); Section 3.3 (Table 4)]
    "we divide the speaking speed into five levels, which include very slow, slow, moderate, fast, and very fast, and these levels are determined by the average phoneme speaking frequency. ... we selected 200 high phoneme rate samples from seed-tts test-zh as fast set and 200 low phoneme rate samples as slow set"

    The subjective condition τin is defined as a quantized function of the average phoneme speaking frequency, which is computed from the duration sequence Dtarget that the model is trained and evaluated to predict (Eq. 3: minθ EΦ||Dtarget − f(Φ, E(τin, τex); θ)||^2). In the speed-control ablation, the fast/slow test subsets are selected using the same ground-truth phoneme rate, showing τin is taken from target durations rather than from an independent estimator. DurFormer therefore receives a coarse summary of the answer during evaluation, while Ratio-Scale and FastSpeech2 do not; the reported 15.05%/11.26% MSE gains and heterogeneous-sample WER gains are then partly attributable to conditioning on target-derived information, not to the duration model alone.

full rationale

DurFormer is a trained regression model, and most of its components (attribute encoder, semantic adapter, Gaussian probability head) are standard conditional predictors trained against MFA ground-truth durations; there is no self-citation chain or imported uniqueness theorem, and the reported MSE arithmetic is straightforward. The one structural circularity risk is the speed-level condition τin: Section 2.2.1 defines τin by average phoneme speaking frequency, i.e., as a function of the target duration sequence Dtarget, and Section 3.3 selects fast/slow test sets using ground-truth phoneme rate. If, as the ablation implies, τin is set from ground-truth target durations during evaluation, DurFormer is compared against baselines that lack access to a target-derived summary, making part of the reported 15.05%/11.26% MSE and WER improvements a by-construction advantage rather than a modeling advance. This is a partial, self-definitional circularity in the evaluation; it does not make the entire model equivalent to its inputs because the model must still predict the full duration sequence from text and semantics.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on MFA-provided ground truth, the Gaussian assumption on durations, the sufficiency of speed and scene as conditions, and the transferability of a pretrained language model to Mandarin rhythm. No code, data, or operational definitions for the speed and scene conditions are given, which is the main reproducibility and validity risk.

free parameters (4)
  • Model weights and hyperparameters of DurFormer = not reported
    All neural network parameters are learned on WenetSpeech4TTS subsets, and no parameter counts, layer sizes, or training schedules are reported.
  • Speaking speed level thresholds = not reported
    The paper divides speed into five levels 'determined by average phoneme speaking frequency' but does not give thresholds or the extraction procedure. This hand-chosen quantization is a free design choice.
  • Speech scene categories = not reported
    Speech scene is described as an objective speaking condition, but the categories and their assignment are not defined, leaving the conditioning input unspecified.
  • Duration time unit = 10 ms
    Durations are measured and predicted in 10 ms units, which sets the scale of the loss and all error metrics.
assumptions (5)
  • domain assumption MFA aligner provides correct phoneme-level ground-truth durations
    All training targets and evaluation references come from Montreal Forced Aligner, as stated in Section 3. If MFA is biased, the measured improvements are relative to that aligner rather than to true durations.
  • domain assumption Phoneme duration sequence follows a multivariate Gaussian with diagonal covariance
    Equation (6) uses a Gaussian negative log-likelihood with mean and variance, assuming the duration vector is normally distributed and conditionally independent across phonemes. Natural speech durations can be multimodal or correlated.
  • domain assumption Speaking speed and speech scene are sufficient external conditions for duration variability
    Section 2.2.1 introduces these two factors as the explicit subjective and objective conditions. Other factors mentioned in the introduction, such as emphasis, emotion, and prosody, are not encoded.
  • domain assumption Pretrained Roberta semantic vectors transfer to Mandarin phoneme rhythm
    Section 2.2.2 uses a pretrained Roberta as semantic extractor with an MLP domain-adaptation layer. The specific model version and language are unspecified, and no validation shows that semantics help beyond the attribute features.
  • standard math Text-to-speech alignment is monotonic and decomposes as duration expansion
    Section 2.1 defines S(Phi) as the duration expansion D ⊙ Phi with sum d_i = T. This is the conventional monotonic alignment assumption used throughout neural TTS.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Duration Model for Text Speech Alignment." pith.science (2026). https://pith.science/paper/KAKXCTOP

@misc{pith2026250722612,
  author       = {Pith},
  title        = {Pith review of: Adaptive Duration Model for Text Speech Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KAKXCTOP}},
  note         = {Machine review of arXiv:2507.22612}
}
read the original abstract

Speech-to-text alignment is a critical component of neural text to speech (TTS) models. Autoregressive TTS models typically use an attention mechanism to learn these alignments on-line, while non-autoregressive end to end TTS models rely on durations extracted from external sources. In this paper, we propose a novel duration prediction framework that can give promising phoneme-level duration distribution with given text. In our experiments, the proposed duration model has more precise prediction and adaptation ability to conditions, compared to previous baseline models. Specifically, it makes a considerable improvement on phoneme-level alignment accuracy and makes the performance of zero-shot TTS models more robust to the mismatch between prompt audio and input audio.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 15 canonical work pages

  1. [1]

    A typical TTS system includes an encoder, a decoder, and an alignment mechanism linking linguistic and acoustic representations [3–6]

    INTRODUCTION Neural text-to-speech (TTS) models, particularly autoregres- sive ones, achieve high naturalness on in-domain text but often suffer from pronunciation errors such as word skipping or repetition when generalizing to long or out-of-domain inputs [1–3]. A typical TTS system includes an encoder, a decoder, and an alignment mechanism linking lingu...

    work page

  2. [2]

    Adaptive Duration Model for Text Speech Alignment

    ALIGNMENT LEARNING FRAMEWORK 2.1. Problem Formulation. Given a text-speech data pair < Text, Speech >, typically the text prompt will be transformed into phonemes and be encoded into latent embeddings Φ = {ϕ1, ..., ϕN }, Φ ∈ RCtext×N , where N represents the length of phonemes, and the speech waveform will be transformed into mel- spectrogram X = {x1, ......

    work page Pith review arXiv 2025

  3. [3]

    We use Premium and Basic subsets of Wenet- Speech4TTS as our experiment dataset

    EXPERIMENTS We evaluate the effectiveness of the alignment learning framework by comparing its performance in terms of dis- tance from annotated ground truth durations and synthesized speech quality. We use Premium and Basic subsets of Wenet- Speech4TTS as our experiment dataset. For ground truth durations, we adopt MFA aligner to obtain phoneme level du-...

    work page 2020

  4. [4]

    Durformer outperforms baseline methods with respect to efficiency and accuracy

    CONCLUSION In this paper, we formulate an optimization problem for text speech alignment, and propose an adaptive duration model named DurFormer to predict the phoneme-level duration sequence. Durformer outperforms baseline methods with respect to efficiency and accuracy. We innovatively propose probability module to enable diversity, which is expected to...

    work page

  5. [5]

    One tts align- ment to rule them all,

    Rohan Badlani, Adrian La ´ncucki, Kevin J Shih, Rafael Valle, Wei Ping, and Bryan Catanzaro, “One tts align- ment to rule them all,” in ICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6092–6096

    work page 2022

  6. [6]

    Spark-tts: An efficient llm-based text-to-speech model with single- stream decoupled speech tokens,

    Xinsheng Wang, Mingqi Jiang, Ziyang Ma, Ziyu Zhang, Songxiang Liu, Linqin Li, Zheng Liang, Qixi Zheng, Rui Wang, Xiaoqin Feng, et al., “Spark-tts: An efficient llm-based text-to-speech model with single- stream decoupled speech tokens,” arXiv preprint arXiv:2503.01710, 2025

    arXiv 2025

  7. [7]

    Flowtron: an autoregressive flow-based generative network for text-to-speech synthesis,

    Rafael Valle, Kevin Shih, Ryan Prenger, and Bryan Catanzaro, “Flowtron: an autoregressive flow-based generative network for text-to-speech synthesis,” arXiv preprint arXiv:2005.05957, 2020

    arXiv 2005

  8. [8]

    Fastspeech: Fast, robust and controllable text to speech,

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

    work page 2019

Show all 25 references
  1. [9]

    Fastpitch: Parallel text-to-speech with pitch prediction,

    Adrian La ´ncucki, “Fastpitch: Parallel text-to-speech with pitch prediction,” in ICASSP 2021-2021 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6588–6592

    2021

  2. [10]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” arXiv preprint arXiv:2006.04558, 2020

    2006 arXiv

  3. [11]

    Location-relative attention mechanisms for ro- bust long-form speech synthesis,

    Eric Battenberg, RJ Skerry-Ryan, Soroosh Mariooryad, Daisy Stanton, David Kao, Matt Shannon, and Tom Bagby, “Location-relative attention mechanisms for ro- bust long-form speech synthesis,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Proces...

    2020

  4. [12]

    Seed-tts: A family of high-quality versatile speech generation models,

    Philip Anastassiou, Jiawei Chen, Jitong Chen, Yuanzhe Chen, Zhuo Chen, Ziyi Chen, Jian Cong, Lelai Deng, Chuang Ding, Lu Gao, et al., “Seed-tts: A family of high-quality versatile speech generation models,” arXiv preprint arXiv:2406.02430, 2024

    2024 arXiv

  5. [13]

    Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,

    Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, De- tai Xin, Dongchao Yang, Yanqing Liu, Yichong Leng, Kaitao Song, Siliang Tang, et al., “Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,” arXiv preprint arXiv:2403.03100 , 2024

    2024 arXiv

  6. [14]

    Non-autoregressive neural text-to-speech,

    Kainan Peng, Wei Ping, Zhao Song, and Kexin Zhao, “Non-autoregressive neural text-to-speech,” in Interna- tional conference on machine learning . PMLR, 2020, pp. 7586–7598

    2020

  7. [15]

    Durian-e 2: Duration informed attention net- work with adaptive variational autoencoder and adver- sarial learning for expressive text-to-speech synthesis,

    Yu Gu, Qiushi Zhu, Guangzhi Lei, Chao Weng, and Dan Su, “Durian-e 2: Duration informed attention net- work with adaptive variational autoencoder and adver- sarial learning for expressive text-to-speech synthesis,” in ICASSP 2024-2024 IEEE International Conference on Acoustics,...

    2024

  8. [16]

    Prosody transfer in neural text to speech using global pitch and loudness features,

    Siddharth Gururani, Kilol Gupta, Dhaval Shah, Zahra Shakeri, and Jervis Pinto, “Prosody transfer in neural text to speech using global pitch and loudness features,” arXiv preprint arXiv:1911.09645, 2019

    1911 arXiv

  9. [17]

    Simple-tts: End-to-end text-to-speech synthesis with latent diffusion,

    Justin Lovelace, Soham Ray, Kwangyoun Kim, Kilian Q Weinberger, and Felix Wu, “Simple-tts: End-to-end text-to-speech synthesis with latent diffusion,” 2023

    2023

  10. [18]

    Neural codec lan- guage models are zero-shot text to speech synthesizers,

    Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al., “Neural codec lan- guage models are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  11. [19]

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,

    Jonathan Shen, Ruoming Pang, Ron J Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, Rj Skerrv-Ryan, et al., “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,” in 2018 IEEE international conference on acousti...

    2018

  12. [20]

    Portaspeech: Portable and high-quality generative text-to-speech,

    Yi Ren, Jinglin Liu, and Zhou Zhao, “Portaspeech: Portable and high-quality generative text-to-speech,” Advances in Neural Information Processing Systems , vol. 34, pp. 13963–13974, 2021

    2021

  13. [21]

    V oicebox: Text-guided multilingual universal speech generation at scale,

    Matthew Le, Apoorv Vyas, Bowen Shi, Brian Kar- rer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al., “V oicebox: Text-guided multilingual universal speech generation at scale,” Advances in neural information processing sys- tems, vol...

    2023

  14. [22]

    Maskgct: Zero-shot text-to-speech with masked generative codec transformer,

    Yuancheng Wang, Haoyue Zhan, Liwei Liu, Rui- hong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu, “Maskgct: Zero-shot text-to-speech with masked generative codec transformer,” arXiv preprint arXiv:2409.00750, 2024

    2024 arXiv

  15. [23]

    Sim- plespeech: Towards simple and efficient text-to-speech with scalar latent transformer diffusion models,

    Dongchao Yang, Dingdong Wang, Haohan Guo, Xueyuan Chen, Xixin Wu, and Helen Meng, “Sim- plespeech: Towards simple and efficient text-to-speech with scalar latent transformer diffusion models,” arXiv preprint arXiv:2406.02328, 2024

    2024 arXiv

  16. [24]

    F5- tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen, “F5- tts: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  17. [25]

    Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model benchmark,

    Linhan Ma, Dake Guo, Kun Song, Yuepeng Jiang, Shuai Wang, Liumeng Xue, Weiming Xu, Huan Zhao, Binbin Zhang, and Lei Xie, “Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model benchmark,” arXiv preprint arXiv:2406.05763, 2024

    2024 arXiv

Pith tools

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