Traceable TTS: Toward Watermark-Free TTS with Strong Traceability

Kai Yu; Shuai Wang; Xie Chen; Yunchong Xiao; Yushen Chen; Yuxiang Zhao; Zhikang Niu

REVIEW 3 major objections 5 minor 44 references

Traceable TTS: Toward Watermark-Free TTS with Strong Traceability

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 Joint training of a TTS model and a paired discriminator can make synthesized speech traceable to its source model without embedding watermarks.

desk verdict Neat cooperative-training idea for watermark-free TTS traceability, but the main experiment conflates finetuning with the joint objective; missing control undercuts the headline claim. read the letter →

arxiv 2507.03887 v1 pith:OBPX44YA submitted 2025-07-05 eess.AS

Yuxiang Zhao , Yunchong Xiao , Yushen Chen , Zhikang Niu , Shuai Wang , Kai Yu , Xie Chen This is my paper

classification eess.AS

keywords watermark-freeTTStraceabilitymodelattributionjointtrainingF5-TTSsynthesizedspeechaudiodeepfakedetectiongenerativefingerprinting

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 proposes a way to trace which text-to-speech (TTS) model produced a given audio clip without embedding a watermark. Instead of adding an explicit n-bit watermark to the waveform or vocoder, it jointly trains the TTS generator and a paired discriminator with aligned objectives: the same binary-cross-entropy loss used by the discriminator is backpropagated into the TTS model so the generator learns to produce audio the discriminator can recognize, and the discriminator learns to recognize that audio. Using F5-TTS as the testbed, the paper reports that the joint discriminator attributes audio to the right source on out-of-domain data with 89.38% accuracy versus 85.60% for the independently trained baseline, while WER, speaker similarity, and UTMOS quality scores improve slightly. If the claim holds, model attribution can work without degrading speech quality or relying on watermarks that can be spoofed, and it continues to work when the vocoder is swapped.

What carries the argument

The mechanism that carries the argument is a generator–discriminator pair trained with the same objective, not opposing ones. The generator is F5-TTS, a flow-matching diffusion TTS that is end-to-end differentiable from waveform back to text, which is what lets the discriminator's BCE loss reach the TTS parameters. The discriminator is a wav2vec 2.0 feature extractor followed by a nine-layer LCNN classifier. The combined loss is $L = \lambda_{\text{TTS}} L_{\text{TTS}} + \lambda_{\text{BCE}} L_{\text{BCE}}$ with both weights set to 1; $L_{\text{BCE}}$ is the same loss for both models. Training runs as a three-stage loop: finetune F5-TTS with the joint loss, reconstruct the LibriTTS dev-clean set by voice cloning so speaker and text content are controlled, then train the discriminator on the reconstructed audio, with original-F5-TTS audio included as hard negatives. The loop repeats ten times, after which the discriminator trains for ten more epochs without those hard negatives.

What would settle it

Train a discriminator with the same architecture and schedule on audio from F5-TTS finetuned for the same number of epochs on LibriTTS train-clean-100 but without any discriminator loss, and evaluate it on the same out-of-domain test set with the same stratified sampling; if its ACC reaches or exceeds 89.38% (or AUC reaches or exceeds 0.9421), the gain is due to finetuning, not to the aligned joint objective.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that reversing the sign of the adversarial objective—making the generator and discriminator cooperate rather than compete—imprints a recognizable, quality-neutral signature onto synthesized speech. During joint training, the discriminator is trained to label jointly finetuned F5-TTS audio as 1, real speech as 0, and audio from the original unfinetuned F5-TTS as 0; the same BCE loss is propagated into F5-TTS's parameters. After ten joint-training loops plus ten further discriminator epochs without the original-F5 negatives, the learned signature transfers to unseen data: out-of-domain AUC rises from 0.8823 to 0.9421 and ACC from 85.60% to 89.38% when test negatives include audio from CosyVoice, CosyVoice 2, and E2-TTS, while in-domain accuracy stays near 99.7–99.9% and the finetuned model's WER, speaker similarity, and UTMOS all improve slightly.

Load-bearing premise

The load-bearing premise is that the out-of-domain accuracy gain comes from the joint-training loop itself, but the only reported comparison pairs independently trained discrimination of original F5-TTS audio against jointly trained discrimination of finetuned F5-TTS audio, leaving no control for finetuning alone.

Editorial extensions

If this is right

A TTS model can be made traceable at training time without any watermark payload, removing watermark-induced quality loss and the risk of forged watermarks.
The signature is tied to the acoustic model rather than the vocoder, so attribution survives replacing the vocoder with a new one.
The learned signature generalizes to audio from unseen TTS systems: out-of-domain accuracy is 89.38% versus 85.60% for the baseline, and AUC rises from 0.8823 to 0.9421.
The method is robust to resampling, speed changes, reverb, volume changes, and MP3/WAV conversion, with additive noise and pitch shift as the main failure points.
The approach requires end-to-end differentiability, so discrete-token TTS models such as VALL-E are not directly supported; the paper leaves a universal training framework for future work.

Reading between the lines

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

The reported comparison does not isolate finetuning from joint training; a control discriminator trained on audio from F5-TTS finetuned for the same 20 epochs without the discriminator loss would separate the two effects, and that control is a direct test of the paper's causal claim.
The equal weights on the TTS and BCE losses are chosen without a sensitivity study, so scaling the BCE term relative to the TTS term could trade traceability against audio quality; the observed slight quality improvement suggests the operating point may not be optimal.
If an adversary obtains the jointly finetuned model, they might fine-tune it against the paired discriminator to erase the learned signature; the paper does not test such adversarial spoofing.
The aligned-objective idea is not restricted to TTS: any differentiable generator paired with a classifier could in principle be trained this way for attribution, so the mechanism may transfer to other generative modalities.

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.

Desk Editor's Note

read the letter

Quick take: the paper proposes a neat idea—training a TTS model and a discriminator with aligned objectives so the discriminator can attribute audio to that specific TTS without watermarking—but the headline result is undercut by a missing control, and the current experiments don't support the causal claim.

What's genuinely new: framing watermark-free TTS traceability as a cooperative, sign-flipped GAN objective is a real departure from both watermarking and passive fingerprinting. The paper is also honest about scope: it only works for differentiable TTS models, and the quality numbers (WER, SIM, UTMOS) suggest the extra loss doesn't hurt audio quality. That's worth something.

The soft spot is the one the stress-test points at, and it's load-bearing. In Table 3, the baseline discriminator was trained on audio from the original F5-TTS; 'ours' was trained on audio from F5-TTS finetuned for 20 epochs with the joint loss. So two things changed at once: the training objective and the generator checkpoint. A plain finetuning-only control (same 20 epochs, lambda_BCE=0) is missing. Without it, the 3.78-point ACC gain cannot be attributed to the cooperative objective. The paper even says in Sec 5.2 that its test uses finetuned F5-TTS audio, so it's not an oversight in the write-up—it's a designed choice that conflates the variables. Also, the final 10 epochs of discriminator training after discarding the original-model negatives is a post-hoc adjustment that changes what the discriminator is asked to do; it's hard to say how much of the gain comes from that.

Minor issues: no error bars or repeated runs, no code or data yet, and the 'strong traceability' claim isn't tested against spoofing or adversarial examples. Robustness to noise and pitch is already shaky (ACC drops ~13% and ~7%), so 'strong' is overstatement.

Who this is for: audio deepfake/forensics people who want an active attribution method for differentiable TTS. The idea is publishable in a better form. As is, the central result is underdetermined. I'd like to see a finetuning-only control, repeated runs with variance, and a spoofing test; then the claim can be evaluated fairly.

Recommendation: send to peer review? Yes—the question is real, the method is novel, and the flaw is identifiable and fixable. But the current version should not be accepted without the control.

Referee Report

3 major / 5 minor

Summary. The paper proposes a watermark-free TTS traceability framework in which an F5-TTS generator and a wav2vec 2.0 + LCNN discriminator are trained jointly with an aligned BCE objective. The stated goal is for the generator to produce speech that the paired discriminator can recognize while preserving or improving audio quality, and for the discriminator to attribute audio to the jointly trained TTS model even on out-of-domain data. Experiments compare a baseline discriminator trained on original F5-TTS audio with the proposed system trained on audio from F5-TTS finetuned for 20 epochs with the joint loss, reporting improved out-of-domain accuracy (89.38% vs 85.60%) and preserved or slightly improved quality (Table 5). The paper claims that the joint training loop significantly improves traceability generalization.

Significance. The task formulation is timely and the idea of co-training a TTS generator with its attribution discriminator is interesting; if validated, it would be a useful watermark-free complement to explicit watermarking, with falsifiable predictions about out-of-domain attribution and robustness. The paper is clearly motivated, uses standard evaluation metrics (AUC, EER, ACC), evaluates against several other TTS models, and reports robustness and speech-quality results. The code release promise is also a positive. However, the central causal claim is currently underdetermined because the comparison in Table 3 changes both the training objective and the generator checkpoint, and the equations in Section 3.1 are inconsistent with the label convention used in Section 3.3. These issues directly affect the main claim and need to be resolved before the result can be accepted.

major comments (3)

[§4.2–§4.3, §5.2, Table 3] The central comparison conflates the joint-training objective with a change of generator checkpoint. The baseline discriminator is trained on audio from the original open-source F5-TTS, while the proposed method first finetunes F5-TTS for 20 epochs (10 loops × 2 epochs) using L = L_TTS + L_BCE, then trains and retrains its discriminator on audio from that finetuned model; Section 5.2 explicitly states that the test also uses audio from the finetuned F5-TTS. Two variables therefore change at once: whether the BCE term is added during finetuning, and which F5-TTS checkpoint produces the positive-class training and test audio. A control discriminator trained on audio from an F5-TTS model finetuned for the same 20 epochs with λ_BCE = 0, and evaluated on that model's audio, is necessary to attribute the 3.78-point ACC gain in Table 3 to the joint objective rather than to ordinary finetuning. Without this control, the Abstract's claim that joint training 'significantly improves traceability generalization' is empirically underdetermined.
[§3.1, Eqs. (3)–(4)] The sign of the generator objective is inconsistent with the label convention in Section 3.3. Equations (3) and (4) are the standard GAN equations: D maximizes log D(x) + log(1−D(G(z))) and G maximizes log(1−D(G(z))), which drives D(G(z)) toward 0. But Section 3.3 assigns label 1 to F5-TTS generated audio and label 0 to ground truth, and states that the goal is for the audio generated by F5-TTS to be 'easier to be detected by the discriminator as true', which requires the generator to increase D(G(z)), i.e., to minimize log(1−D(G(z))) or maximize log D(G(z)). As written, the equations describe a generator that makes D reject its own output, which is the opposite of the described joint objective. This is a load-bearing inconsistency because the method is defined by this joint objective; the equations should be corrected or the label semantics clarified.
[§5.2, Table 1 and Table 3] The out-of-domain evaluation is asymmetric in the positive class. The baseline is tested on original-F5-TTS audio, while the proposed method is tested on finetuned-F5-TTS audio, so the two systems are never evaluated on the same positive-class distribution. This makes it impossible to separate the effect of the joint objective from the effect of the changed generator distribution. A complete comparison should cross-evaluate both discriminators on both original-F5-TTS and finetuned-F5-TTS audio, using the same negative-class data. Section 5.2's phrase 'to ensure a fair comparison' is misleading; the current design does not provide a controlled comparison.

minor comments (5)

[§3.1, Eq. (4)] The text says the generator and discriminator are assigned 'the same loss function', but Eq. (4) contains only the generator term while Eq. (3) contains both the real and generated terms; the relation between the two objectives should be stated precisely.
[§4.2–§4.3] The paper does not report confidence intervals or significance tests for the accuracy differences in Tables 2 and 3; given that the main result is a 3.78-point ACC gap, some measure of variability would help assess its reliability.
[§5.4, Table 3] The text claims 'consistently higher accuracy rates across various test datasets', but Table 3 reports only aggregate AUC, EER, and ACC; per-category accuracy for CosyVoice, CosyVoice 2, and E2-TTS would support this claim.
[References] BigVGAN (Ref. [14]) and MUSAN (Ref. [24]) are listed in the bibliography but not cited in the text; the placeholder '[?]' should be replaced with the proper citations.
[Table 4] The column header 'W A V' should read 'WAV'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the joint-training loop is the proposed method, and the traceability evaluation is externally grounded against held-out speakers and other TTS models. The missing finetuning-only control is a confound, not a circular derivation.

full rationale

The paper's central claim is empirical rather than derivational: joint training of F5-TTS and a paired discriminator improves out-of-domain traceability. The generator and discriminator are indeed trained in a closed loop by design (Sec. 3.3, Eq. 5), but that is the method being proposed, not a hidden circular step. The evaluation is not self-referential in a vacuous sense: the discriminator is tested on held-out speakers and text from LibriTTS test-clean and must distinguish the paired generator's audio from ground truth and from three other TTS models (CosyVoice, CosyVoice 2, E2-TTS), which provides external grounding. The self-citation of F5-TTS [4] overlaps with the author list, but F5-TTS is the model under study rather than evidence used to justify the traceability result. The one substantive weakness is an experimental confound: the baseline discriminator is trained on audio from the original F5-TTS checkpoint (Sec. 4.2), while the proposed method is trained on audio from F5-TTS finetuned for 20 epochs under the joint objective (Sec. 4.3), and Sec. 5.2 explicitly states that the test positive class also uses the finetuned model. Without a control discriminator trained on audio from a finetuned-only F5-TTS (with the BCE term disabled), the improvement in Table 3 cannot be causally attributed to the joint-training objective. This is a missing control and a correctness risk, not a circular reduction: the reported accuracy is not equal to any fitted parameter or training label by construction. Therefore no circularity is found, and the score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The listed hyperparameters are all hand-set; none are fitted to the test set. The missing ablation for the number of loops and loss weights means the method's sensitivity is unknown. The paper introduces no new physical entities, mediators, or architectural components beyond a standard classifier; the paired discriminator is a known architecture (wav2vec 2.0 + LCNN).

free parameters (6)

lambda_TTS = 1
Hand-set loss weight for the F5-TTS term in Eq. 5; no sensitivity analysis provided.
lambda_BCE = 1
Hand-set loss weight for the discriminator BCE term in Eq. 5; no sensitivity analysis provided.
joint_training_loops = 10
Number of finetune-generate-train cycles in Sec 4.3; chosen without ablation or stopping criterion.
finetune_epochs_per_loop = 2
Epochs of F5-TTS finetuning per loop in Sec 4.3; no rationale given for this value.
discriminator_epochs = 20 per loop, plus 10 final
Discriminator training schedule in Sec 4.3; the final 10 epochs are added after discarding the original-F5-TTS negatives.
learning_rate = 5e-6
Adam learning rate for discriminator training (Sec 4.2); standard but hand-set.

assumptions (5)

domain assumption F5-TTS is differentiable end-to-end from waveform back to text, so the discriminator's BCE loss can be backpropagated through the frozen vocoder into the acoustic model.
Required by the joint training in Sec 3.3, Eq. 5; noted as a limitation in Sec 6 because token-based TTS models like VALL-E are excluded.
domain assumption The frozen pre-trained vocoder (Vocos or BigVGAN) is differentiable enough to propagate gradients without collapsing training.
The finetuning phase trains only the acoustic model, so gradients must pass through the vocoder (Sec 3.3); this is not verified in the paper.
domain assumption The wav2vec 2.0 + LCNN architecture is a suitable discriminator for traceability of unseen speakers and text.
Chosen based on prior deepfake detection results (Sec 3.2); the paper does not compare discriminator architectures.
ad hoc to paper Ten loops of the finetune-generate-train cycle are sufficient to converge to a useful trace.
Sec 4.3 sets 10 loops without ablation or a stopping criterion.
domain assumption Balanced stratified sampling in the test phase yields a fair and reliable evaluation.
Sec 4.4 describes down-sampling to balance classes; the paper does not report the resulting sample sizes per condition.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Traceable TTS: Toward Watermark-Free TTS with Strong Traceability." pith.science (2026). https://pith.science/paper/OBPX44YA

@misc{pith2026250703887,
  author       = {Pith},
  title        = {Pith review of: Traceable TTS: Toward Watermark-Free TTS with Strong Traceability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBPX44YA}},
  note         = {Machine review of arXiv:2507.03887}
}

read the original abstract

Recent advances in Text-To-Speech (TTS) technology have enabled synthetic speech to mimic human voices with remarkable realism, raising significant security concerns. This underscores the need for traceable TTS models-systems capable of tracing their synthesized speech without compromising quality or security. However, existing methods predominantly rely on explicit watermarking on speech or on vocoder, which degrades speech quality and is vulnerable to spoofing. To address these limitations, we propose a novel framework for model attribution. Instead of embedding watermarks, we train the TTS model and discriminator using a joint training method that significantly improves traceability generalization while preserving-and even slightly improving-audio quality. This is the first work toward watermark-free TTS with strong traceability. To promote progress in related fields, we will release the code upon acceptance of the paper.

Figures

Figures reproduced from arXiv: 2507.03887 by the authors.

**Figure 2.** Overall framework of the proposed method. Unlike GAN which makes synthetic and real data indistinguishable, our [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Model architecture of TTS model and discriminator [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Receiver Operating Characteristic (ROC) compari [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 23 canonical work pages

[1]

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

work page 2020
[2]

Alimohammad Beigi, Zhen Tan, Nivedh Mudiam, Canyu Chen, Kai Shu, and Huan Liu. 2024. Model attribution in llm-generated disinformation: A domain generalization approach with supervised contrastive learning. In Proc. DSAA. IEEE

work page 2024
[3]

Guangyu Chen, Yu Wu, Shujie Liu, Tao Liu, Xiaoyong Du, and Furu Wei. 2023. Wavmark: Watermarking for audio generation. arXiv preprint arXiv:2308.12770 (2023)

arXiv 2023
[4]

Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. 2024. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. arXiv preprint arXiv:2410.06885 (2024). Traceable TTS: Toward Watermark-Free TTS with Strong Traceability

arXiv 2024
[5]

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

arXiv 2024
[6]

Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, et al . 2024. Cosyvoice 2: Scal- able streaming speech synthesis with large language models. arXiv preprint arXiv:2412.10117 (2024)

arXiv 2024
[7]

Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. 2019. Gltr: Statis- tical detection and visualization of generated text.arXiv preprint arXiv:1906.04043 (2019)

arXiv 2019
[8]

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (2020)

work page 2020

Show all 44 references

[9]

Haorui He, Zengqiang Shang, Chaoren Wang, Xuyuan Li, Yicheng Gu, Hua Hua, Liwei Liu, Chen Yang, Jiaqi Li, Peiyang Shi, et al . 2024. Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation. In Proc. SLT. IEEE

2024
[10]

Syou Hirofumi, Kazuto Fukuchi, Yohei Akimoto, and Jun Sakuma. 2022. Did you use my gan to generate fake? Post-hoc attribution of gan generated images via latent recovery. In Proc. IJCNN. IEEE

2022
[11]

Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processin...

2021
[12]

Jee-weon Jung, Hee-Soo Heo, Ju-ho Kim, Hye-jin Shim, and Ha-Jin Yu. 2019. Rawnet: Advanced end-to-end deep neural network using raw waveforms for text-independent speaker verification. arXiv preprint arXiv:1904.08104 (2019)

2019 arXiv
[13]

Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu, and Nicholas Evans. 2022. AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks. In Proc. ICASSP. IEEE

2022
[14]

Sang-gil Lee, Wei Ping, Boris Ginsburg, Bryan Catanzaro, and Sungroh Yoon
[15]

Linyang Li, Pengyu Wang, Ke Ren, Tianxiang Sun, and Xipeng Qiu. 2023. Origin tracing and detecting of llms. arXiv preprint arXiv:2304.14072 (2023)

2023 arXiv
[16]

Weizhi Liu, Yue Li, Dongdong Lin, Hui Tian, and Haizhou Li. 2024. GROOT: Generating robust watermark for diffusion-model-based audio synthesis. In Proc. ACM MM

2024
[17]

Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. 2019. Do gans leave artificial fingerprints?. In Proc. MIPR. IEEE

2019
[18]

Aleksandr Meister, Matvei Novikov, Nikolay Karpov, Evelina Bakhturina, Vitaly Lavrukhin, and Boris Ginsburg. 2023. LibriSpeech-PC: Benchmark for evaluation of punctuation and capitalization capabilities of end-to-end asr models. In Proc. ASRU. IEEE

2023
[19]

Mahieyin Rahmun, Rafat Hasan Khan, Tanjim Taharat Aurpa, Sadia Khan, Zulker Nayeen Nahiyan, Mir Sayad Bin Almas, Rakibul Hasan Rajib, and Syeda Sakira Hassan. 2024. Synthetic speech classification: IEEE signal pro- cessing cup 2022 challenge. arXiv preprint arXiv:2412.13279 (2024)

2024 arXiv
[20]

Yong Ren, Jiangyan Yi, Tao Wang, Jianhua Tao, Zhengqi Wen, Chenxing Li, Zheng Lian, Ruibo Fu, Ye Bai, and Xiaohui Zhang. 2025. P2Mark: Plug-and-play parameter-intrinsic watermarking for neural speech generation. arXiv preprint arXiv:2504.05197 (2025)

2025 arXiv
[21]

Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. 2022. UTMOS: Utokyo-sarulab system for voicemos challenge 2022. arXiv preprint arXiv:2204.02152 (2022)

2022 arXiv
[22]

Yuhui Shi, Qiang Sheng, Juan Cao, Hao Mi, Beizhe Hu, and Danding Wang. 2024. Ten words only still help: Improving black-box AI-generated text detection via proxy-guided efficient re-sampling. arXiv preprint arXiv:2402.09199 (2024)

2024 arXiv
[23]

Hubert Siuzdak. 2023. Vocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis. arXiv preprint arXiv:2306.00814 (2023)

2023 arXiv
[24]

David Snyder, Guoguo Chen, and Daniel Povey. 2015. Musan: A music, speech, and noise corpus. arXiv preprint arXiv:1510.08484 (2015)

2015 arXiv
[25]

Rafael Rivera Soto, Kailin Koch, Aleem Khan, Barry Chen, Marcus Bishop, and Nicholas Andrews. 2024. Few-shot detection of machine-generated text using style representations. arXiv preprint arXiv:2401.06712 (2024)

2024 arXiv
[26]

Xu Tan, Tao Qin, Frank Soong, and Tie-Yan Liu. 2021. A survey on neural speech synthesis. arXiv preprint arXiv:2106.15561 (2021)

2021 arXiv
[27]

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

2023 arXiv
[28]

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

2025 arXiv
[29]

Kangxi Wu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-Seng Chua. 2023. LLMDet: A third party large language models generated text detection tool.arXiv preprint arXiv:2305.15004 (2023)

2023 arXiv
[30]

Xinrui Yan, Jiangyan Yi, Jianhua Tao, Chenglong Wang, Haoxin Ma, Tao Wang, Shiming Wang, and Ruibo Fu. 2022. An initial investigation for detecting vocoder fingerprints of fake audio. In Proceedings of the 1st International Workshop on Deepfake Detection for Audio Multimedia

2022
[31]

Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. 2022. Deepfake network architecture attribution. In Proc. AAAI, Vol. 36

2022
[32]

Tianyun Yang, Danding Wang, Fan Tang, Xinying Zhao, Juan Cao, and Sheng Tang. 2023. Progressive open space expansion for open-set model attribution. In Proc. CVPR

2023
[33]

Xianjun Yang, Wei Cheng, Yue Wu, Linda Petzold, William Yang Wang, and Haifeng Chen. 2023. Dna-gpt: Divergent n-gram analysis for training-free detec- tion of gpt-generated text. arXiv preprint arXiv:2305.17359 (2023)

2023 arXiv
[34]

Jiangyan Yi, Chu Yuan Zhang, Jianhua Tao, Chenglong Wang, Xinrui Yan, Yong Ren, Hao Gu, and Junzuo Zhou. 2024. ADD 2023: Towards audio deepfake detection and analysis in the wild. arXiv preprint arXiv:2408.04967 (2024)

2024 arXiv
[35]

Ning Yu, Larry S Davis, and Mario Fritz. 2019. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proc. ICCV

2019
[36]

Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. 2021. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proc. ICCV

2021
[37]

Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. 2020. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726 (2020)

2020 arXiv
[38]

Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. 2019. Libritts: A corpus derived from librispeech for text-to- speech. arXiv preprint arXiv:1904.02882 (2019)

2019 arXiv
[39]

Chu Yuan Zhang, Jiangyan Yi, Jianhua Tao, Chenglong Wang, and Xinrui Yan
[40]

Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. 2023. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137 (2023)

2023 arXiv
[41]

Junzuo Zhou, Jiangyan Yi, Yong Ren, Jianhua Tao, Tao Wang, and Chu Yuan Zhang. 2025. WMCodec: End-to-end neural speech codec with deep watermark- ing for authenticity verification. In Proc. ICASSP

2025
[42]

Junzuo Zhou, Jiangyan Yi, Tao Wang, Jianhua Tao, Ye Bai, Chu Yuan Zhang, Yong Ren, and Zhengqi Wen. 2024. Traceablespeech: Towards proactively traceable text-to-speech with watermarking. arXiv preprint arXiv:2406.04840 (2024)

2024 arXiv
[2022]

arXiv preprint arXiv:2206.04658 (2022)

Bigvgan: A universal neural vocoder with large-scale training. arXiv preprint arXiv:2206.04658 (2022)

2022 arXiv
[2024]

In China National Conference on Chinese Computational Linguistics

Distinguishing neural speech synthesis models through fingerprints in speech waveforms. In China National Conference on Chinese Computational Linguistics. Springer

Pith tools

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

[1] [1]

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

work page 2020

[2] [2]

Alimohammad Beigi, Zhen Tan, Nivedh Mudiam, Canyu Chen, Kai Shu, and Huan Liu. 2024. Model attribution in llm-generated disinformation: A domain generalization approach with supervised contrastive learning. In Proc. DSAA. IEEE

work page 2024

[3] [3]

Guangyu Chen, Yu Wu, Shujie Liu, Tao Liu, Xiaoyong Du, and Furu Wei. 2023. Wavmark: Watermarking for audio generation. arXiv preprint arXiv:2308.12770 (2023)

arXiv 2023

[4] [4]

Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. 2024. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. arXiv preprint arXiv:2410.06885 (2024). Traceable TTS: Toward Watermark-Free TTS with Strong Traceability

arXiv 2024

[5] [5]

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

arXiv 2024

[6] [6]

Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, et al . 2024. Cosyvoice 2: Scal- able streaming speech synthesis with large language models. arXiv preprint arXiv:2412.10117 (2024)

arXiv 2024

[7] [7]

Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. 2019. Gltr: Statis- tical detection and visualization of generated text.arXiv preprint arXiv:1906.04043 (2019)

arXiv 2019

[8] [8]

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (2020)

work page 2020