NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation

Berrak Sisman; Haizhou Li; Huhong Xian; Rui Liu

REVIEW 4 major objections 7 minor 24 references

NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation

T0 review · 4 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read NE-PADD claims that fusing named-entity attention into frame-level deepfake detectors cuts equal error rate from 11.59% to 7.89% on PartialSpoof-NER.

desk verdict NE-PADD is a plausible new idea—semantic NER attention for PADD—but the reported gains are not yet shown to come from named-entity semantics rather than from an extra attention branch. read the letter →

arxiv 2509.03829 v1 pith:I6A455JZ submitted 2025-09-04 cs.CL

Huhong Xian , Rui Liu , Berrak Sisman , Haizhou Li This is my paper

classification cs.CL

keywords partialaudiodeepfakedetectionnamedentityrecognitionattentionfusiontransferframe-levelspooflocalizationSpoof-NERequalerrorrate

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

NE-PADD is a frame-level partial audio deepfake detector that tries to use named-entity semantics as a detection cue. It runs a frozen speech named-entity recognition branch in parallel with a standard partial-spoof detection branch and merges their attention signals, either by gated feature fusion (AF) or by a KL-divergence transfer loss (AT). On a new PartialSpoof-NER dataset, built by transcribing PartialSpoof with an ASR system and tagging entities automatically, the fusion variant reports 7.89% equal error rate, down from 11.59% for the strongest prior baseline. The same fusion and transfer modules also improve two existing detectors, BAM and TDL, suggesting the semantic cue is portable rather than tied to one architecture.

What carries the argument

The load-bearing object is the pair of frame-level attention distributions, α_ADD from the PADD branch and α_NER from the frozen SpeechNER branch. α_NER, learned by a speech NER encoder operating on log-spectrograms, marks where named entities occur; α_ADD marks where spoofed frames are suspected. AF fuses the attended hidden states with a learned gate g = σ(W_g[H_ADD-attended, H_NER-attended]) and feeds the convex combination to the classifier. AT instead minimizes asymmetric KL divergence between the two attention distributions, letting α_NER act as teacher. Together these mechanisms are the entire route through which entity semantics reach the frame-level classifier.

What would settle it

Run the same architecture with the entity-attention weights replaced by random or shuffled attention; if the equal error rate stays near 7.89%, named-entity semantics are not the cause. Separately, measure the overlap between entity spans and the spoofed-frame labels in PartialSpoof-NER; if overlap is near chance, the assumed semantic alignment is unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, NE-PADD establishes that injecting named-entity attention into a frame-level deepfake detector materially improves partial spoof localization. The PADD branch extracts Wav2Vec2.0 features through residual blocks and produces attention weights α_ADD; the frozen SpeechNER branch, built on an end-to-end English speech NER encoder, produces α_NER. Attention Fusion computes a sigmoid gate over the concatenated attended embeddings to blend them, while Attention Transfer treats α_NER as a teacher and adds KL(α_NER∥α_ADD) to the binary cross-entropy loss. The best configuration, NE-PADD-AF, reaches 7.89% EER on PartialSpoof-NER, outperforming WBD's 11.59%; adding the same m

Load-bearing premise

The gain rests on the frozen speech-entity recognizer actually marking the frames that matter for spoofing; the paper reports neither the recognizer's accuracy nor the measured overlap between entity spans and spoofed frames.

Editorial extensions

If this is right

Frame-level partial deepfake detection can be improved by adding a frozen semantic branch; the PADD branch itself need not be retrained from scratch to benefit.
Named-entity attention transfers across architectures: the same AF and AT modules improve BAM and TDL, not just the proposed NE-PADD backbone.
PartialSpoof-NER becomes a benchmark resource annotated with 11,572, 1,407, and 1,526 named entities across train, dev, and eval splits, enabling future semantic PADD work.
The 7.89% EER on PartialSpoof-NER is the reference point that subsequent partial deepfake detectors will be measured against.

Reading between the lines

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

The causal role of named entities is not yet isolated: replacing α_NER with random-but-stable attention in the same architecture would show how much of the 3.7-point gain is semantic alignment versus auxiliary multi-task regularization.
The authors do not report NER accuracy or the overlap between entity spans and spoofed frames; if that overlap is low, the benefit would have to come from some other property of the NER branch's features.
Because AF outperforms AT in the head-to-head comparison, gated feature fusion appears to carry more of the gain than distribution matching; a combined AF+AT variant is the natural next configuration to test.
On datasets whose partial forgeries target non-entity words, numbers, or prosodic units rather than names, the semantic shortcut would weaken; the method's scope is likely tied to entity-centric manipulation.

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

The short version: this paper does something new—coupling a frozen SpeechNER attention branch with a frame-level partial audio deepfake detector—and the strongest evidence is that the same attention aggregation (fusion or transfer) improves not just their own model but also two published baselines, BAM and TDL. That transfer suggests the mechanism is doing something useful, not just memorizing a benchmark.

What's genuinely new: first combination of SpeechNER with PADD via attention fusion/transfer, and the PartialSpoof-NER dataset built from PartialSpoof with Whisper-based transcription and Stanza NER annotations. Code and data are promised. That is a useful resource even if the headline effect shrinks under closer inspection.

The main soft spot is exactly what the stress-test note flags: there is no non-semantic control branch. The SpeechNER branch is frozen, but it still adds parameters, a gating mechanism, and an auxiliary loss—any of which could improve a frame classifier without conveying named-entity meaning. Nothing in the paper shows that alpha_NER is actually identifying entities on this data; there is no NER F1 on PartialSpoof-NER, no alignment accuracy, no overlap statistics between named entity spans and spoofed frames. So the claim 'named entity knowledge improves PADD' is underdetermined. A simple control—same architecture, but with random or untrained attention in the NER branch—would settle it.

Beyond that, the reporting choices hedge the headline number: they pick the better of two variants (AF over AT) as 'optimal,' tune lambda_KL per model, report no repeated runs or confidence intervals, and the baselines look like published numbers rather than re-runs under matched conditions. Those are real but modest concerns for a workshop-level paper; they don't undermine the observation that attention aggregation helps.

The paper is clearly written, the method is simple and reproducible, and the dataset construction is transparent. It is a reasonable first step in a new direction. My recommendation: send it to peer review, and ask the authors to add the non-semantic control and report NER quality/overlap before the semantic claim is accepted. I'd cite it for the dataset and the attention-aggregation idea, but I wouldn't yet cite it as evidence that entity semantics is what matters.

Referee Report

4 major / 7 minor

Summary. The paper proposes NE-PADD, a frame-level partial audio deepfake detection (PADD) method that augments a Wav2Vec2.0/ResNet1D backbone with a frozen SpeechNER branch. Two attention aggregation mechanisms are introduced: Attention Fusion (AF), which gates and combines PADD and NER attended features (Eqs. 7-8), and Attention Transfer (AT), which regularizes the PADD attention with KL divergence against the NER attention (Eqs. 9-10). The authors construct a new dataset, PartialSpoof-NER, by annotating PartialSpoof audio with Whisper/Stanza named-entity spans. On this dataset, the AF variant is reported to achieve 7.89% EER, outperforming the best baseline WBD at 11.59%, and applying AF/AT to BAM and TDL also improves their EER. The paper also reports a finer-grained experiment showing lower EER as the number of spoofed segments increases.

Significance. If the central claim holds, the paper would demonstrate that named-entity semantic information, injected through attention, materially benefits partial audio deepfake detection, a task where prior work has largely ignored semantics. The new PartialSpoof-NER dataset and the public code are useful community assets. However, the current experimental design does not isolate the semantic content of the auxiliary NER branch from the generic effect of adding an extra attention branch or auxiliary regularization, so the main contribution is underdetermined as presented. The selection of the better of two variants and per-model tuning of lambda_KL further weaken the headline comparison.

major comments (4)

[Sections II-C, II-E, III-E (Table II)] The experiments do not include a control branch that isolates the semantic content of the SpeechNER attention. NE-PADD-AF and NE-PADD-AT add an extra attention branch, a gating mechanism, and/or an auxiliary KL loss; any of these could improve frame-level classification as a regularizer or as a second acoustic view, independent of named-entity semantics. To support the title-level claim that named-entity knowledge is the source of the gain, add a control in which the frozen attention labels are randomized or the NER branch is replaced by a same-capacity branch trained without semantic labels. Also report NER accuracy/span F1 on PartialSpoof-NER and the overlap between NER spans and spoofed frames. Without this evidence, Table II's 7.89% vs. 11.59% gap is ambiguous.
[Section III-E, Table II, Figure 2] The reported representative result uses NE-PADD-AF because it 'delivers the best empirical results' (Section III-E), and for AT variants lambda_KL is tuned per model across 0.1-1.0 (Section III-F, Figure 2). This post-hoc selection, combined with no error bars or repeated-seed variance, makes the headline 7.89% and the ablation gains in Table III hard to interpret. Report both AF and AT without best-of-two selection, give means/standard deviations over multiple runs, and specify the lambda selection procedure (e.g., chosen on the adaptation set before eval).
[Section III-B and Table II] It is unclear whether the baseline EERs in Table II were produced by re-running the baselines under the same conditions on PartialSpoof-NER, or whether they are copied from the original papers. Since PartialSpoof-NER retains the same audio as PartialSpoof, using previously reported numbers may be defensible, but the evaluation protocol, train/adapt/eval split, and frame-level scoring must be stated explicitly. The same applies to the BAM-AT/BAM-AF and TDL-AT/TDL-AF variants in Table III: specify common settings (backbone initialization, optimizer, epochs, and the adaptation-set early stopping).
[Section II-C and III-A] The SpeechNER feature extraction module is a central component, yet its implementation details are missing. The paper does not state whether a pretrained SpeechNER checkpoint is used, which dataset it was trained on, how the character-level NE tags are aligned to the 20 ms frame grid used by Wav2Vec2.0, or what frame-level NER accuracy is achieved. Without this information, the reliability of the alpha_NER teacher signal in Eq. (5) and its use in Eq. (9) cannot be assessed. Please provide the checkpoint source, training recipe, and alignment procedure, and report NER quality on the evaluation set.

minor comments (7)

[Abstract] 'Speech Name Entity Recognition' should be 'Speech Named Entity Recognition' for consistency with the acronym SpeechNER.
[Section II-B] Typo: 'Wav2Vec2.0' appears in the text; use 'Wav2Vec 2.0' consistently.
[Table II] The baseline name 'Single reso.' is cryptic; spell out the full method name or reference number.
[Eq. (9)] The KL divergence is written over i and j, but the attention distributions alpha_NER and alpha_ADD are not explicitly defined as per-sequence or per-frame distributions. Clarify the summation ranges so that Eq. (9) is dimensionally consistent.
[Table I] The column headers 'Bona fide', 'Fake', 'All' and 'Named Entities Count' are not fully defined (utterances vs. segments). Clarify what each count represents.
[Figures 2 and 3] Figure 2's x-axis is labeled with raw values but no axis title; Figure 3's y-axis jumps to 65 for forgery level 1 but the surrounding text only describes trend. Add axis titles and discuss the spike.
[Section III-G] The finer-grained experiment is said to be limited by dataset size, but no quantitative statement of subset sizes is given. Report the number of utterances per forgery-level subset.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the NER attention signal comes from a frozen branch, the PADD labels are not used to define it, and the headline EER is an empirical comparison against baselines.

full rationale

The derivation chain is not circular. The PADD branch (Eqs. 3-4) and the SpeechNER branch (Eqs. 5-6) are separate; α_NER is produced by a frozen SpeechNER module and is not defined in terms of the frame-level spoof labels. The fusion (Eqs. 7-8) is a trainable gated combination, not an identity, and the attention-transfer loss (Eqs. 9-10) is an auxiliary regularizer, not a redefinition of the detection target. The claimed result (EER 7.89% in Table II) is a measured evaluation-set number, not a quantity forced by the objective. The paper transparently reports that NE-PADD-AF was chosen as the best of the two variants ('Since NE-PADD-AF delivers the best empirical results, we report its performance as the representative implementation'), and λ_KL tuning in Fig. 2 is standard hyperparameter selection rather than a fitted-value-renamed-as-prediction. There are no load-bearing self-citations or imported uniqueness theorems. The lack of a non-semantic control branch is a real experimental-design limitation—it leaves open whether the gain comes from named-entity semantics or from any auxiliary attention branch—but that is an underdetermination/validity concern, not a circularity, because the central claim does not reduce to its inputs by construction. The acknowledged limitations (short spoofed segments, lack of Chinese-English data) are stated openly and do not indicate circular reasoning.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No physical entities; the central free choice is the transfer weight lambda_KL. The model uses standard learned parameters (the gating weight g), which are not counted. The main added assumptions are about the reliability and relevance of the NER signal.

free parameters (1)

lambda_KL (attention transfer weight) = selected per model by sweeping 0.1-1.0; not reported
Used in the KL transfer loss (Eq. 10) and tuned per model according to Fig. 2 and Table III.

assumptions (3)

domain assumption Spoofed segments in partial deepfakes frequently contain named entities.
Motivates the entire NE-PADD design (Section I); never quantitatively demonstrated on PartialSpoof or PartialSpoof-NER, no overlap statistics provided.
domain assumption Whisper ASR transcriptions and Stanza NER annotations of PartialSpoof audio are accurate enough to provide reliable named-entity supervision.
Used to construct PartialSpoof-NER (Section III-A); no NER quality metrics reported.
domain assumption The frozen SpeechNER encoder's attention (alpha_NER) provides frame-level semantic signal aligned with Wav2Vec2 features.
Attention transfer/fusion in Sections II-D and II-E operate on alpha_NER and H_NER assuming 20ms frame correspondence; training data and alignment are not described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation." pith.science (2026). https://pith.science/paper/I6A455JZ

@misc{pith2026250903829,
  author       = {Pith},
  title        = {Pith review of: NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6A455JZ}},
  note         = {Machine review of arXiv:2509.03829}
}

read the original abstract

Different from traditional sentence-level audio deepfake detection (ADD), partial audio deepfake detection (PADD) requires frame-level positioning of the location of fake speech. While some progress has been made in this area, leveraging semantic information from audio, especially named entities, remains an underexplored aspect. To this end, we propose NE-PADD, a novel method for Partial Audio Deepfake Detection (PADD) that leverages named entity knowledge through two parallel branches: Speech Name Entity Recognition (SpeechNER) and PADD. The approach incorporates two attention aggregation mechanisms: Attention Fusion (AF) for combining attention weights and Attention Transfer (AT) for guiding PADD with named entity semantics using an auxiliary loss. Built on the PartialSpoof-NER dataset, experiments show our method outperforms existing baselines, proving the effectiveness of integrating named entity knowledge in PADD. The code is available at https://github.com/AI-S2-Lab/NE-PADD.

Figures

Figures reproduced from arXiv: 2509.03829 by the authors.

**Figure 1.** The architecture of our proposed model. The model architecture includes two feature extraction modules on the left, with the proposed attention fusion [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** EER Results for Different Values of λKL [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** EER results for the dataset with different levels of segmentation [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

[1]

Diffcss: Diverse and expressive conversational speech synthesis with diffusion models,

W. Wu, Z. Lin, Y . Zhou, et al. , “Diffcss: Diverse and expressive conversational speech synthesis with diffusion models,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025
[2]

Stable- tts: Stable speaker-adaptive text-to-speech synthesis via prosody prompting,

W. Han, M. Kang, C. Kim, and E. Yang, “Stable- tts: Stable speaker-adaptive text-to-speech synthesis via prosody prompting,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025
[3]

Grad-stylespeech: Any-speaker adaptive text-to-speech synthesis with dif- fusion models,

M. Kang, D. Min, and S. J. Hwang, “Grad-stylespeech: Any-speaker adaptive text-to-speech synthesis with dif- fusion models,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2023, pp. 1–5

work page 2023
[4]

Freevc: Towards high- quality text-free one-shot voice conversion,

J. Li, W. Tu, and L. Xiao, “Freevc: Towards high- quality text-free one-shot voice conversion,” in IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

work page 2023
[5]

Speechsplit2.0: Unsupervised speech disen- tanglement for voice conversion without tuning autoen- coder bottlenecks,

C. H. Chan, K. Qian, Y . Zhang, and M. Hasegawa- Johnson, “Speechsplit2.0: Unsupervised speech disen- tanglement for voice conversion without tuning autoen- coder bottlenecks,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 6332–6336

work page 2022
[6]

Again-vc: A one-shot voice conversion using activation guidance and adaptive instance normalization,

Y . Chen, D. Wu, T. Wu, and H. Lee, “Again-vc: A one-shot voice conversion using activation guidance and adaptive instance normalization,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5954–5958

work page 2021
[7]

Avqvc: One-shot voice conversion by vector quantiza- tion with applying contrastive learning,

H. Tang, X. Zhang, J. Wang, N. Cheng, and J. Xiao, “Avqvc: One-shot voice conversion by vector quantiza- tion with applying contrastive learning,” in IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 4613–4617

work page 2022
[8]

Enhancing expressive voice conversion with discrete pitch-conditioned flow matching model,

J. Zuo, S. Ji, M. Fang, et al. , “Enhancing expressive voice conversion with discrete pitch-conditioned flow matching model,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025

Show all 24 references

[9]

Hifi-gan: Generative ad- versarial networks for efficient and high fidelity speech synthesis,

J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative ad- versarial networks for efficient and high fidelity speech synthesis,” in Advances in Neural Information Process- ing Systems (NIPS) , vol. 33, 2020, pp. 17 022–17 033

2020
[10]

Wavenet: A generative model for raw audio,

A. Van Den Oord, S. Dieleman, H. Zen, et al. , “Wavenet: A generative model for raw audio,” arXiv preprint arXiv:1609.03499, vol. 12, p. 1, 2016

2016 arXiv
[11]

Natural tts synthesis by conditioning wavenet on mel spectro- gram predictions,

J. Shen, R. Pang, R. J. Weiss, et al. , “Natural tts synthesis by conditioning wavenet on mel spectro- gram predictions,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2018, pp. 4779–4783

2018
[12]

End-to-end anti-spoofing with rawnet2,

H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2021, pp. 6369– 6373

2021
[13]

Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

J.-w. Jung, H.-S. Heo, H. Tak, et al. , “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” in IEEE international conference on acoustics, speech and signal processing (ICASSP) , IEEE, 2022, pp. 6367–6371

2022
[14]

Half-truth: A par- tially fake audio detection dataset,

J. Yi, Y . Bai, J. Tao, et al. , “Half-truth: A par- tially fake audio detection dataset,” arXiv preprint arXiv:2104.03617, 2021

2021 arXiv
[15]

An initial investigation for detecting par- tially spoofed audio,

L. Zhang, X. Wang, E. Cooper, J. Yamagishi, J. Patino, and N. Evans, “An initial investigation for detecting par- tially spoofed audio,” arXiv preprint arXiv:2104.02518, 2021

2021 arXiv
[16]

Multi-task learning in utterance-level and segmental- level spoof detection,

L. Zhang, X. Wang, E. Cooper, and J. Yamagishi, “Multi-task learning in utterance-level and segmental- level spoof detection,”arXiv preprint arXiv:2107.14132, 2021

2021 arXiv
[17]

Wav2vec 2.0: A framework for self-supervised learn- ing of speech representations,

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A framework for self-supervised learn- ing of speech representations,” in Advances in Neural Information Processing Systems (NIPS) , H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33, Curran...

2020
[18]

The partialspoof database and countermeasures for the detection of short fake speech segments embedded in an utterance,

L. Zhang, X. Wang, E. Cooper, N. Evans, and J. Yamag- ishi, “The partialspoof database and countermeasures for the detection of short fake speech segments embedded in an utterance,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 813–825, 2022

2022
[19]

Waveform boundary detection for partially spoofed audio,

Z. Cai, W. Wang, and M. Li, “Waveform boundary detection for partially spoofed audio,” in IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1–5

2023
[20]

End-to-end named entity recognition from english speech,

H. Yadav, S. Ghosh, Y . Yu, and R. R. Shah, “End-to-end named entity recognition from english speech,” arXiv preprint arXiv:2005.11184, 2020

2005 arXiv
[21]

Robust speech recogni- tion via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recogni- tion via large-scale weak supervision,” in International conference on machine learning (ICML) , PMLR, 2023, pp. 28 492–28 518

2023
[22]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, et al., “Attention is all you need,” Advances in neural information process- ing systems (NIPS) , vol. 30, 2017

2017
[23]

Enhancing partially spoofed audio localization with boundary-aware attention mech- anism,

J. Zhong, B. Li, and J. Yi, “Enhancing partially spoofed audio localization with boundary-aware attention mech- anism,” arXiv preprint arXiv:2407.21611 , 2024

2024 arXiv
[24]

An efficient temporary deepfake location approach based embed- dings for partially spoofed audio detection,

Y . Xie, H. Cheng, Y . Wang, and L. Ye, “An efficient temporary deepfake location approach based embed- dings for partially spoofed audio detection,” in IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP) , IEEE, 2024, pp. 966–970

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Diffcss: Diverse and expressive conversational speech synthesis with diffusion models,

W. Wu, Z. Lin, Y . Zhou, et al. , “Diffcss: Diverse and expressive conversational speech synthesis with diffusion models,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025

[2] [2]

Stable- tts: Stable speaker-adaptive text-to-speech synthesis via prosody prompting,

W. Han, M. Kang, C. Kim, and E. Yang, “Stable- tts: Stable speaker-adaptive text-to-speech synthesis via prosody prompting,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025

[3] [3]

Grad-stylespeech: Any-speaker adaptive text-to-speech synthesis with dif- fusion models,

M. Kang, D. Min, and S. J. Hwang, “Grad-stylespeech: Any-speaker adaptive text-to-speech synthesis with dif- fusion models,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2023, pp. 1–5

work page 2023

[4] [4]

Freevc: Towards high- quality text-free one-shot voice conversion,

J. Li, W. Tu, and L. Xiao, “Freevc: Towards high- quality text-free one-shot voice conversion,” in IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

work page 2023

[5] [5]

Speechsplit2.0: Unsupervised speech disen- tanglement for voice conversion without tuning autoen- coder bottlenecks,

C. H. Chan, K. Qian, Y . Zhang, and M. Hasegawa- Johnson, “Speechsplit2.0: Unsupervised speech disen- tanglement for voice conversion without tuning autoen- coder bottlenecks,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 6332–6336

work page 2022

[6] [6]

Again-vc: A one-shot voice conversion using activation guidance and adaptive instance normalization,

Y . Chen, D. Wu, T. Wu, and H. Lee, “Again-vc: A one-shot voice conversion using activation guidance and adaptive instance normalization,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5954–5958

work page 2021

[7] [7]

Avqvc: One-shot voice conversion by vector quantiza- tion with applying contrastive learning,

H. Tang, X. Zhang, J. Wang, N. Cheng, and J. Xiao, “Avqvc: One-shot voice conversion by vector quantiza- tion with applying contrastive learning,” in IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 4613–4617

work page 2022

[8] [8]

Enhancing expressive voice conversion with discrete pitch-conditioned flow matching model,

J. Zuo, S. Ji, M. Fang, et al. , “Enhancing expressive voice conversion with discrete pitch-conditioned flow matching model,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2025, pp. 1–5

work page 2025