Pith. sign in

REVIEW 5 major objections 6 minor 34 references

Investigation of Zero-shot Text-to-Speech Models for Enhancing Short-Utterance Speaker Verification

T0 review · 5 major / 6 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 Zero-shot TTS-generated speech, fused at the embedding level with the original short utterance, reduces the equal error rate by 10%–16% relative in speaker verification without retraining the ASV model.

desk verdict Useful first look at test-time ZS-TTS for speaker verification, but the headline 10–16% gain is an oracle selection, not a fixed-protocol result. read the letter →

arxiv 2506.14226 v1 pith:DMQLOJO4 submitted 2025-06-17 cs.SD

Yiyang Zhao , Shuai Wang , Guangzhi Sun , Zehua Chen , Chao Zhang , Mingxing Xu , Thomas Fang Zheng This is my paper · ORCID
classification cs.SD
keywords short-utterancespeakerverificationzero-shottext-to-speechtest-timedataaugmentationembeddingfusionequalerrorrateECAPA-TDNNVoxCeleb1
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 asks whether modern zero-shot text-to-speech models, which can clone a voice from a few seconds of audio, can help when a speaker-verification system is given only a very short test utterance. The authors propose a test-time augmentation that synthesizes a longer utterance from the short one, extracts a speaker embedding from that synthetic speech, and fuses it with the embedding of the original real speech. On VoxCeleb1 with an ECAPA-TDNN verifier, they report 10%–16% relative equal error rate reductions across all tested durations without retraining the verifier, with the largest relative gains for 0.5–1 second speech. The paper also finds that directly replacing real speech with synthetic speech hurts, that text content's phoneme coverage matters, and that longer synthetic speech does not match the benefit of longer real speech.

What carries the argument

The load-bearing mechanism is test-time embedding fusion. The short real utterance $x_b$ is used as the prompt for a zero-shot TTS model to synthesize a longer utterance $x_g$; both are passed through the same frozen ECAPA-TDNN encoder to obtain embeddings $e_b$ and $e_g$, which are then combined by weighted mean, concatenation, addition, or a gated attention fusion. The fusion is the step that turns imperfect synthetic speech into a measurable gain: direct replacement degrades performance, but combining the embeddings yields complementary speaker information, and the gain is not explained by duration alone since repeating the short utterance to the same length does not reproduce it.

What would settle it

Run the same fusion protocol but fix the fusion weight from a development partition (or use a default like $w=0.5$) and generate several synthetic samples per utterance to estimate variance; if the relative EER reduction over baseline drops below 10% or becomes non-significant, the claimed benefit does not generalize. A second check is to retrain the ASV model or test on VoxCeleb2; if the gains disappear, the effect is specific to the ECAPA-TDNN/VoxCeleb1 setup.

Watch

Extended reading notes

Core claim

The central claim is that zero-shot TTS-generated speech, when used as a test-time complement rather than a replacement, carries speaker-discriminative information that a fixed speaker-verification model can exploit. The concrete result is that fusing the original embedding $e_b$ with the embedding $e_g$ of TTS-generated speech using a weighted mean $w e_b + (1-w) e_g$ or a learned gated attention fusion reduces EER by 10%–16% relative over the baseline on VoxCeleb1, with the largest relative improvements at 0.5s and 1s. The best reported result combines CosyVoice and MaskGCT embeddings in a two-stage fusion, reaching 13.0%, 16.0%, 16.7%, and 3.6% relative EER reductions at 0.5s, 1s, 2s, and full-length segments respectively at an optimal fusion weight of $w=0.6$.

Load-bearing premise

The headline 10%–16% reduction relies on the fusion weight being chosen by searching on the same test set used for reporting, so the claim assumes that weight would transfer to a fixed deployment setting and that the single reported EER is stable across different randomly generated TTS samples.

Editorial extensions

If this is right

  • Test-time TTS augmentation is a drop-in enhancement for existing verification systems: no retraining, no score normalization, and no change to the enrollment pipeline are required.
  • The benefit concentrates on very short speech (0.5–1s), where relative EER reductions are largest, making the method most relevant for access-control, forensic, and mobile scenarios with brief utterances.
  • Direct substitution is harmful; any deployment should fuse, not replace, the real and synthetic embeddings.
  • Fusing multiple TTS models (CosyVoice and MaskGCT) gives the best and most weight-stable results, suggesting that diversity across synthesis architectures is valuable.
  • Synthetic speech does not fully substitute for real duration: longer real speech remains better, so TTS augmentation is a partial remedy rather than a replacement for collecting longer utterances.

Reading between the lines

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

  • The test-set-swept fusion weight means the reported 10%–16% range is likely an upper bound; a fixed-weight or development-tuned protocol would probably shrink the gains, and the paper does not report variance across TTS generation seeds.
  • The phoneme-coverage result (EER rising from 8.63% to 13.73% as phoneme count drops from 42 to 7) suggests the text prompt itself can be optimized per utterance, for example by using the original transcript or a text planner that maximizes phoneme diversity.
  • The same fusion idea could extend beyond speech to other biometrics where a generative model can produce a longer or cleaner probe from a short enrollment sample, such as face or gait recognition.
  • Because the fusion weight is stable across durations when two TTS models are combined, a practical system might fix $w$ near 0.6 and skip per-condition tuning.
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

5 major / 6 minor

Summary. The paper proposes a test-time data augmentation pipeline for short-utterance speaker verification: a zero-shot TTS model generates a longer speech sample from each short bona fide utterance, the ECAPA-TDNN speaker encoder embeds both the original and the synthetic speech, and the two speaker embeddings are fused by addition, concatenation, weighted mean, or a learned attention gate. The authors evaluate three pre-trained ZS-TTS systems (NaturalSpeech 3, CosyVoice, MaskGCT) on VoxCeleb1 across four test durations and report that fusing real and synthetic embeddings yields 10%–16% relative EER reductions across durations, with the largest relative gains for 0.5–1 s speech, without retraining the ASV model. The paper also analyzes how TTS model choice, target text/phoneme coverage, speech duration, and fusion strategy affect performance, and reports a two-stage CosyVoice+MaskGCT combination as the best configuration.

Significance. If the reported gains are confirmed under a fixed evaluation protocol, this would be a useful and novel contribution: it is, to my knowledge, the first systematic study of using zero-shot TTS for test-time augmentation in speaker verification, and it shows that synthetic speech cannot simply replace real speech but can complement it through embedding fusion. The paper is methodologically transparent about the TTS systems and fusion strategies it compares, and it reports a number of useful negative results, such as the degradation when using TTS output alone and the reduced benefit of longer synthetic speech relative to longer real speech. The experiments use standard public resources (VoxCeleb1, ECAPA-TDNN via WeSpeaker), which aids reproducibility in principle. However, the central quantitative claim is currently weakened by evaluation-protocol issues: the fusion weight appears to be selected on the same test set used for reporting, no variance or repeated-run information is given, and the best-performing configuration is selected from many alternatives. These issues are fixable but must be addressed before the headline result can be accepted.

major comments (5)
  1. [§4.4 and Figure 3] The fusion weight w in the weighted-mean fusion (Eq. 1) is selected by sweeping on the VoxCeleb1 test set and reporting the EER at the best w, as shown in Figure 3 and described in Section 4.4. This makes the reported weighted-mean results an oracle estimate rather than a fixed-protocol estimate. The paper must either specify a pre-chosen weight, choose w on a development split, or report the sensitivity of the EER to w (e.g., the EER at a single fixed w such as w=0.5) for all conditions in Table 1.
  2. [Abstract and Table 1] The abstract's claim of '10%–16% relative EER reductions across all durations' is not supported by a single fixed configuration in Table 1. For example, the CosyVoice+MaskGCT weighted-mean row with the reported w=0.6 gives relative reductions of 14.3% at 0.5 s, 16.0% at 1 s, and 16.7% at 2 s, but only about 4.9% at full length; individual TTS models show even smaller or negligible full-length reductions (e.g., NaturalSpeech 3 weighted mean gives 2.05% at full length, matching the baseline). The text in Section 4.4 reports reductions for a fixed w=0.6 separately from the best-per-condition numbers, and the paper should clearly separate these two types of results and state which numbers support the abstract.
  3. [Eqs. (2)–(3) and §2.2] The attention-based fusion method trains a gated fusion network with a speaker classifier and loss, but the manuscript does not state what data are used for this training. If this network is trained or fine-tuned on the VoxCeleb1 test set (or on the same utterances used for evaluation), the reported EERs for the attention fusion condition are not a valid estimate of performance on unseen data. Please specify the training/validation split, the optimization details, and the number of parameters updated.
  4. [Tables 1 and 3 (experimental results)] Each experimental condition is reported as a single EER value with no confidence intervals, no standard errors across TTS synthesis seeds, and no repeated evaluation trials. Since ZS-TTS generation is stochastic and the reported differences are small in absolute terms (e.g., MaskGCT weighted mean at 0.5 s, 25.97% vs. 26.83% baseline; CosyVoice weighted mean at full length, 2.02% vs. 2.05% baseline), it is unclear whether the observed improvements are stable. The paper should report results over multiple TTS sampling runs or provide confidence intervals for the key comparisons.
  5. [§3 and Footnote 1] All conclusions are based on a single dataset (VoxCeleb1) and a single ASV model (ECAPA-TDNN), as Footnote 1 explicitly acknowledges. The paper's central claim of 10%–16% relative EER reductions across durations is presented as a general property of the proposed method, but the evidence is limited to one data domain and one backbone. At minimum, the abstract and conclusion should be rephrased to state the scope, and ideally the method should be validated on a second dataset or a second speaker encoder to show that the fusion gains are not an artifact of the particular ECAPA-TDNN checkpoint.
minor comments (6)
  1. [Throughout] There are frequent typographical spacing issues in the text, such as 'V oxCeleb', 'V oxCeleb 1', and 'CosyV oice' in the abstract and body; these should be corrected to 'VoxCeleb1' and 'CosyVoice' for consistency.
  2. [§4.3] The phrase 'less then 2 seconds' should be 'less than 2 seconds'.
  3. [Figure 1(b)] The notation in the merge-method diagram is unclear: the label '1-W' appears without a definition, and the diagram does not clearly indicate which branch is weighted by w and which by 1−w. Please align the figure labels with Eq. (1).
  4. [Table 1] The column header 'All' is ambiguous; the text refers to 'full-length' utterances, and the table should use the same term to avoid confusion between 'all durations' and the full-length condition.
  5. [§4.2 and Table 2] The phoneme-coverage experiment is conducted only with CosyVoice and only at 2 s prompts; the paper should state this limitation in the same paragraph, since it is not clear that the phoneme effect generalizes to other TTS models or shorter prompts.
  6. [§4.4] The final paragraph mentions fusing three generated utterances from the same prompt, but no quantitative results are reported for this multi-utterance fusion condition; please either provide the numbers or mark it as a direction for future work.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 10–16% EER reductions are obtained by sweeping the fusion weight on the same VoxCeleb1 test set used for reporting, making the headline gain an in-sample optimum rather than a fixed-protocol result.

  1. fitted input called prediction [Abstract; Section 4.4, Table 1, Figure 3]
    "Figure 3 caption: 'Comparison of EER Performance for Different TTS Models with Varying Origin Embedding Weights across Different Speech Durations (Fusion with weighted mean)'; Section 4.4: 'For instance, with an optimal fusion weight of w= 0.6, the relative EER reductions are 3.6%, 16.7%, 16.0%, and 13.0% for full-length, 2s, 1s, and 0.5s segments, respectively.'"

    The abstract's headline claim that 'combining real and synthetic speech samples leads to 10%-16% relative equal error rate (EER) reductions across all durations' is computed after selecting the fusion weight w on the same VoxCeleb1 test set used to report the EERs. Figure 3 explicitly sweeps w across durations and TTS models, and Section 4.4 reports the result at the 'optimal fusion weight.' Thus the reported reduction is the minimum of the test-set EER surface over w (and, more broadly, the best among several TTS models, fusion strategies, and text-selection policies in Table 1), not the performance of a pre-specified fixed protocol.

full rationale

This paper contains no mathematical derivation chain; it is an experimental study. The main circularity-like issue is not self-citation or definitional equivalence but test-set optimization presented as a headline result. The reported EER reductions depend on the fusion weight w being swept on the very VoxCeleb1 test set used for evaluation, as shown in Figure 3 and Section 4.4. Consequently, the abstract's 10–16% relative improvement range is the value of the optimization objective at the selected optimum, not an out-of-sample estimate. The paper does not fix the fusion weight before evaluation, does not validate on a held-out set, and does not report variance across TTS generations. These facts undermine the quantitative claim but do not invalidate the qualitative direction that embedding fusion of ZS-TTS-generated speech can help short-utterance verification. There are no load-bearing self-citations: reference [6] is a prior paper by the same authors but is not used to justify the central result, and reference [9] is a general overview. The central statistical issue is therefore a fitted-input-called-prediction problem, warranting a score of 6 rather than a higher definitional-circularity score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method is empirical and rests on the validity of the VoxCeleb1 truncated trial protocol, the speaker-preservation capability of zero-shot TTS from short prompts, and the assumption that embedding-space averaging is a sound fusion operation. The paper explicitly notes in footnote 1 that conclusions are restricted to VoxCeleb1. The main free parameter is the fusion weight w, tuned on the test set, which directly affects the reported gains.

free parameters (2)
  • Fusion weight w in weighted mean fusion (Eq. 1) = 0.6 for the CosyVoice+MaskGCT two-stage fusion; per-model optimal values vary across durations (Figure 3)
    Selected on the VoxCeleb1 test set to minimize EER; the reported gains are a result of this test-set optimization, not a predetermined weight. Section 4.4 and Figure 3 show the weight sweep.
  • Attention fusion network parameters (Eq. 2-3) = Learned, not specified
    Trained with the speaker classifier loss in Section 2.2; the training split and any interaction with the test set are not stated.
assumptions (5)
  • domain assumption Weighted mean and gated attention of two speaker embeddings yield a valid speaker representation for cosine scoring.
    Equations (1)-(3) in Section 2.2 assume embedding-space arithmetic preserves speaker identity; no theoretical derivation is given.
  • domain assumption Zero-shot TTS models preserve speaker identity from prompts as short as 0.5-2s well enough that generated longer utterances add speaker information.
    Core premise of Section 2.1; the paper itself notes TTS stability declines for short prompts in Section 4.3.
  • domain assumption Truncating each VoxCeleb1 test utterance at its midpoint yields a valid short-utterance evaluation protocol.
    Section 3.1 describes this protocol; it differs from standard VoxCeleb1 trial construction and is not validated against other short-utterance benchmarks.
  • ad hoc to paper A single EER per condition is a stable estimate despite stochastic TTS generation and test-set weight selection.
    Tables 1-2 report one EER per condition with no confidence intervals or repeated runs; the three fixed-text averages in Section 2.1 are not reported as variance.
  • domain assumption VoxCeleb1 is representative for the general claim of enhancing short-utterance speaker verification.
    Footnote 1 restricts conclusions to VoxCeleb1 and notes it is not entirely clean; generalization to other datasets and ASV models is assumed but not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigation of Zero-shot Text-to-Speech Models for Enhancing Short-Utterance Speaker Verification." pith.science (2026). https://pith.science/paper/DMQLOJO4

@misc{pith2026250614226,
  author       = {Pith},
  title        = {Pith review of: Investigation of Zero-shot Text-to-Speech Models for Enhancing Short-Utterance Speaker Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMQLOJO4}},
  note         = {Machine review of arXiv:2506.14226}
}
read the original abstract

Short-utterance speaker verification presents significant challenges due to the limited information in brief speech segments, which can undermine accuracy and reliability. Recently, zero-shot text-to-speech (ZS-TTS) systems have made considerable progress in preserving speaker identity. In this study, we explore, for the first time, the use of ZS-TTS systems for test-time data augmentation for speaker verification. We evaluate three state-of-the-art pre-trained ZS-TTS systems, NatureSpeech 3, CosyVoice, and MaskGCT, on the VoxCeleb 1 dataset. Our experimental results show that combining real and synthetic speech samples leads to 10%-16% relative equal error rate (EER) reductions across all durations, with particularly notable improvements for short utterances, all without retraining any existing systems. However, our analysis reveals that longer synthetic speech does not yield the same benefits as longer real speech in reducing EERs. These findings highlight the potential and challenges of using ZS-TTS for test-time speaker verification, offering insights for future research.

Figures

Figures reproduced from arXiv: 2506.14226 by the authors.

Figure 1
Figure 1. TTS-Enhanced Short-Utterance Speaker Recognition Pipeline. 2. Method In this section, we describe our overall design, which lever￾ages TTS-generated speech to provide additional information for short-utterance speaker recognition, without modifying the original speaker model in any way [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Relative EER Reduction (%) for Different TTS Models across Various Durations (Fusion with weighted mean) phoneme representation in synthetic speech for improved per￾formance. Furthermore, we evaluated two text selection strategies: us￾ing the original transcript for each utterance versus maintain￾ing a consistent synthesized text across all test speech. Results across three TTS models indicate that text uniformity h… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages

  1. [1]

    Investigation of Zero-shot Text-to-Speech Models for Enhancing Short-Utterance Speaker Verification

    Introduction Speaker verification (SV) aims to confirm an individual’s iden- tity based on their voice characteristics. In recent years, deep neural network (DNN)-based models [1–8] have driven remark- able advancements in SV systems. Despite these achievements, SV systems still encounter challenges in short-utterance sce- narios [9]. Unlike longer speech...

    work page Pith review arXiv 2025

  2. [2]

    Figure 1 illustrates the entire workflow along with implementation details

    Method In this section, we describe our overall design, which lever- ages TTS-generated speech to provide additional information for short-utterance speaker recognition, without modifying the original speaker model in any way. Figure 1 illustrates the entire workflow along with implementation details. 2.1. Speaker Embedding Extraction We begin by using th...

    work page

  3. [3]

    Experimental Setup 3.1. Dataset To investigate the effectiveness of the proposed approach, we conduct experiments on the V oxCeleb1 [26] dataset, which contains over 150,000 utterances from 1,251 celebrities. This dataset spans diverse ethnicities and accents, making it a chal- lenging benchmark for speaker recognition. Since our objective is to evaluate ...

    work page

  4. [4]

    In this chap- ter, we will analyze the results from four aspects: TTS models, speech durations, text content, and fusion techniques

    Experimental Results Table 1 presents the overall experimental results. In this chap- ter, we will analyze the results from four aspects: TTS models, speech durations, text content, and fusion techniques. Table 1:EER Performance Comparison Across TTS Architec- tures and Fusion Strategies (Evaluated using ECAPA-TDNN on VoxCeleb1) TTS Model Fusion Method EE...

    work page

  5. [5]

    Conclusions This paper investigates the potential benefits of zero-shot TTS (ZS-TTS) for short-utterance speaker verification. Through ex- periments on TTS models, speech durations, text content, and fusion techniques, we observe that while ZS-TTS provides mea- surable improvements, it also has notable limitations. Directly replacing original speech with ...

    work page

  6. [6]

    X-vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-vectors: Robust dnn embeddings for speaker recognition,” inProc. ICASSP, Calgary, 2018, pp. 5329–5333

    work page 2018

  7. [7]

    Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    arXiv 2005

  8. [8]

    But system description to voxceleb speaker recognition chal- lenge 2019,

    H. Zeinali, S. Wang, A. Silnova, P. Mat ˇejka, and O. Plchot, “But system description to voxceleb speaker recognition chal- lenge 2019,”arXiv preprint arXiv:1910.12592, 2019

    arXiv 2019

Show all 34 references
  1. [9]

    Mfa-conformer: Multi-scale feature aggregation con- former for automatic speaker verification,

    Y . Zhang, Z. Lv, H. Wu, S. Zhang, P. Hu, Z. Wu, H.-y. Lee, and H. Meng, “Mfa-conformer: Multi-scale feature aggregation con- former for automatic speaker verification,” inProc. Interspeech, Inchon, 2022, pp. 306–310

    2022

  2. [10]

    Memory storable network based feature aggregation for speaker representation learning,

    B. Gu, W. Guo, and J. Zhang, “Memory storable network based feature aggregation for speaker representation learning,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 643–655, 2023

    2023

  3. [11]

    Whisper-pmfa: Partial multi-scale feature aggregation for speaker verification using whisper models,

    Y . Zhao, S. Wang, G. Sun, Z. Chen, C. Zhang, M. Xu, and T. F. Zheng, “Whisper-pmfa: Partial multi-scale feature aggregation for speaker verification using whisper models,” inProc. Inter- speech, Kos, 2024, pp. 2680–2684

    2024

  4. [12]

    Cam++: A fast and efficient network for speaker verification using context- aware masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “Cam++: A fast and efficient network for speaker verification using context- aware masking,” inProc. Interspeech, Dublin, 2023, pp. 5301– 5305

    2023

  5. [13]

    Eres2netv2: Boosting short-duration speaker verifica- tion performance with computational efficiency,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, Q. Chen, S. Zhang, and J. Li, “Eres2netv2: Boosting short-duration speaker verifica- tion performance with computational efficiency,”arXiv preprint arXiv:2406.02167, 2024

    2024 arXiv

  6. [14]

    Overview of speaker modeling and its applications: From the lens of deep speaker representation learning,

    S. Wang, Z. Chen, K. A. Lee, Y . Qian, and H. Li, “Overview of speaker modeling and its applications: From the lens of deep speaker representation learning,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, pp. 4971–4998, 2024

    2024

  7. [15]

    A deep neural network for short- segment speaker recognition,

    A. Hajavi and A. Etemad, “A deep neural network for short- segment speaker recognition,”arXiv preprint arXiv:1907.10420, 2019

    1907 arXiv

  8. [16]

    Deep speaker embedding learning with multi-level pooling for text-independent speaker verification,

    Y . Tang, G. Ding, J. Huang, X. He, and B. Zhou, “Deep speaker embedding learning with multi-level pooling for text-independent speaker verification,” inICASSP 2019-2019 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2019, pp. 6116–6120

    2019

  9. [17]

    Improving multi-scale aggregation using feature pyramid module for robust speaker verification of variable-duration utterances,

    Y . Jung, S. M. Kye, Y . Choi, M. Jung, and H. Kim, “Improving multi-scale aggregation using feature pyramid module for robust speaker verification of variable-duration utterances,” inProc. In- terspeech, Shanghai, 2020, pp. 1501–1505

    2020

  10. [18]

    Improving aggregation and loss function for better embedding learning in end-to-end speaker verification system

    Z. Gao, Y . Song, I. McLoughlin, P. Li, Y . Jiang, and L.-R. Dai, “Improving aggregation and loss function for better embedding learning in end-to-end speaker verification system.” inProc. In- terspeech, Graz, 2019, pp. 361–365

    2019

  11. [19]

    Meta- learning for short utterance speaker recognition with imbalance length pairs,

    S. M. Kye, Y . Jung, H. B. Lee, S. J. Hwang, and H. Kim, “Meta- learning for short utterance speaker recognition with imbalance length pairs,” inProc. Interspeech, Shanghai, 2020, pp. 2982– 2986

    2020

  12. [20]

    Text-independent speaker verification with adversarial learning on short utterances,

    K. Liu and H. Zhou, “Text-independent speaker verification with adversarial learning on short utterances,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6569–6573

    2020

  13. [21]

    Short utterance compensation in speaker verification via cosine-based teacher- student learning of speaker embeddings,

    J.-w. Jung, H.-S. Heo, H.-j. Shim, and H.-J. Yu, “Short utterance compensation in speaker verification via cosine-based teacher- student learning of speaker embeddings,” in2019 IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019, pp. 335–341

    2019

  14. [22]

    Open-set short utterance forensic speaker verification using teacher-student network with explicit inductive bias,

    M. Sang, W. Xia, and J. H. Hansen, “Open-set short utterance forensic speaker verification using teacher-student network with explicit inductive bias,” inProc. Interspeech, Shanghai, 2020, pp. 2262–2266

    2020

  15. [23]

    Cnn-based joint mapping of short and long utterance i-vectors for speaker verification us- ing short utterances

    J. Guo, U. A. Nookala, and A. Alwan, “Cnn-based joint mapping of short and long utterance i-vectors for speaker verification us- ing short utterances.” inProc. Interspeech, Stockholm, 2017, pp. 3712–3716

    2017

  16. [24]

    I-vector transformation us- ing conditional generative adversarial networks for short utterance speaker verification,

    J. Zhang, N. Inoue, and K. Shinoda, “I-vector transformation us- ing conditional generative adversarial networks for short utterance speaker verification,” inProc. Interspeech, Hyderabad, 2018, pp. 3613–3617

    2018

  17. [25]

    Data augmentation using deep generative models for embedding based speaker recog- nition,

    S. Wang, Y . Yang, Z. Wu, Y . Qian, and K. Yu, “Data augmentation using deep generative models for embedding based speaker recog- nition,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 28, pp. 2598–2609, 2020

    2020

  18. [26]

    Explor- ing voice conversion based data augmentation in text-dependent speaker verification,

    X. Qin, Y . Yang, L. Yang, X. Wang, J. Wang, and M. Li, “Explor- ing voice conversion based data augmentation in text-dependent speaker verification,”arXiv preprint arXiv:2011.10710, 2020

    2011 arXiv

  19. [27]

    Synaug: Synthesis- based data augmentation for text-dependent speaker verification,

    C. Du, B. Han, S. Wang, Y . Qian, and K. Yu, “Synaug: Synthesis- based data augmentation for text-dependent speaker verification,” inICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5844–5848

    2021

  20. [28]

    Neural codec language mod- els are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Liet al., “Neural codec language mod- els are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  21. [29]

    Cosyvoice: A scalable multi- lingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Maet al., “Cosyvoice: A scalable multi- lingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,”arXiv preprint arXiv:2407.05407, 2024

    2024 arXiv

  22. [30]

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

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  23. [31]

    V oxceleb: Large-scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxceleb: Large-scale speaker verification in the wild,”Computer Speech & Language, vol. 60, p. 101027, 2020

    2020

  24. [32]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” inProc. ICASSP, Rhodes, 2023

    2023

  25. [33]

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

    Y . Wang, H. Zhan, L. Liu, R. Zeng, H. Guo, J. Zheng, Q. Zhang, X. Zhang, S. Zhang, and Z. Wu, “Maskgct: Zero-shot text-to- speech with masked generative codec transformer,”arXiv preprint arXiv:2409.00750, 2024

    2024 arXiv

  26. [34]

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

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin, D. Yang, E. Liu, Y . Leng, K. Song, S. Tang, Z. Wu, T. Qin, X. Li, W. Ye, S. Zhang, J. Bian, L. He, J. Li, and S. Zhao, “Naturalspeech 3: Zero-shot speech syn- thesis with factorized codec and diffusion models,” inForty-first Internati...

    2024

Pith tools

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