Pith. sign in

REVIEW 3 major objections 4 minor 80 references

Controllable Accent Normalization via Discrete Diffusion

T0 review · 3 major / 4 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Accent strength can be dialed by reusing source speech tokens that a predictor scores as already native-like, while discrete diffusion regenerates the rest.

desk verdict Clean systems paper: discrete diffusion + CTP token reuse gives real accent-strength control and the best WER of the compared systems, with the usual small-set/TTS-data caveats. read the letter →

arxiv 2603.14275 v2 pith:P3AEW5ED submitted 2026-03-15 eess.AS cs.AIcs.SD

Qibing Bai , Yuhan Du , Tom Ko , Shuai Wang , Yannan Wang , Haizhou Li This is my paper
classification eess.AScs.AIcs.SD
keywords accentconversionnormalizationdiscretediffusionlanguagemodelself-supervisedspeechtokenscontrollabilitydurationcontrolvoice
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

Most accent-normalization systems force a full shift from non-native to native speech; applications such as language learning and dubbing often need a tunable middle ground. This paper presents DLM-AN, which converts accented speech by running masked discrete diffusion over self-supervised speech tokens. A Common Token Predictor scores each source token by how likely it is to appear in a native rendition; high-scoring tokens can be kept when the diffusion process starts, so reusing more tokens preserves more of the original accent. A separate flow-matching predictor can also stretch or compress total duration to better match native rhythm. On multi-accent English tests the system records the lowest word error rate among the systems compared, competitive accent reduction, and smooth control of accent strength as the reuse threshold is varied.

What carries the argument

Common Token Predictor (CTP): a sequence tagger whose labels are derived from longest-common-subsequence alignment between source and target token sequences; its confidence scores decide which source tokens are reused to initialize the masked discrete-diffusion decoder, thereby controlling accent strength.

What would settle it

On a held-out accent pair whose forced alignments show almost no token overlap even in native-sounding segments, sweep the CTP reuse threshold; if raising the threshold no longer monotonically increases measured accentedness and speaker similarity while decreasing native-likeness, the claimed control mechanism fails.

Watch

Extended reading notes

Core claim

DLM-AN is a controllable accent-normalization system built on masked discrete diffusion over self-supervised speech tokens. A Common Token Predictor identifies source tokens that are likely shared with a native target; selectively reusing those high-confidence tokens to initialize reverse diffusion supplies a simple accent-strength knob—more reuse keeps more of the original accent. A flow-matching Duration Ratio Predictor further sets total duration so the output can follow native rhythm. Experiments on multi-accent English show the lowest word error rate among compared systems together with competitive naturalness and accent reduction, and smooth, interpretable control via the reuse thresho

Load-bearing premise

The method assumes that a phonetically rich tokenizer makes native and accented versions of the same sentence share many tokens exactly in the regions that should be kept, so the predictor’s scores truly mark native-like pronunciation.

Editorial extensions

If this is right

  • A single confidence threshold continuously trades accent reduction against accent retention and speaker similarity.
  • Content preservation (word error rate) stays stronger than the autoregressive token and continuous flow-matching baselines under both free and source-matched duration.
  • Total duration can be scaled independently of accent strength, supporting speaking-rate control for learning or dubbing.
  • The same iterative unmasking procedure supports localized speech infilling or correction.
  • Partial token reuse improves perceived speaker similarity because accent and identity remain entangled.

Reading between the lines

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

  • The same “keep what is already good” reuse idea could transfer to other discrete-token speech edits such as prosody, emotion, or dialect conversion.
  • If LCS-derived CTP labels become noisy for heavily accented speech, a supervised accent-region detector might give a cleaner control surface.
  • Controllable partial normalization may be more useful for pronunciation training than full conversion, because learners often need intermediate targets rather than a finished native model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes DLM-AN, a controllable accent-normalization system that applies masked discrete diffusion (LLaDA-style) over WavLM K-means tokens. A CTC-guided token encoder supplies content features to a Common Token Predictor (CTP), a flow-matching Duration Ratio Predictor, and a bidirectional DLM decoder. CTP is trained with binary labels obtained from longest-common-subsequence alignment of source and target tokens; high-confidence source tokens can be reused to initialize reverse diffusion, giving a continuous accent-strength knob (threshold τ). A separately trained flow-matching synthesizer recovers Mel spectrograms conditioned on speaker embeddings. On an 80-sentence multi-accent L2-ARCTIC test set the system reports the lowest WER among TokAN and CosyAccent baselines, competitive naturalness and accent reduction, monotonic control of ACT/SIM/ΔPPG with τ, and robust duration scaling.

Significance. If the empirical claims hold, the work supplies the first discrete-diffusion formulation of accent normalization that simultaneously offers (i) competitive content preservation and accent reduction and (ii) an interpretable, continuous accent-strength control mechanism via selective token reuse. The combination of masked diffusion, LCS-derived common-token labels, and a duration-ratio predictor is a concrete methodological advance over prior one-shot or continuous-diffusion AC systems, and the progressive ACT/SIM curves with τ are directly useful for language-learning and dubbing applications. Strengths that raise the contribution above a pure engineering paper include the explicit ELBO-style training objective, the CTP-versus-random ablation (Fig. 5), the monotonic τ curves (Fig. 6), and the duration-scaling comparison (Fig. 7).

major comments (3)
  1. §4.1 and §4.4: both training targets and the reference used for ΔPPG are generated by the same native-only Matcha-TTS model. While WER is measured with an external ASR and therefore remains independent, the accent-reduction claim rests partly on similarity to these synthetic references. A small human-native or held-out real-native reference subset (or an explicit statement that ΔPPG is only a proxy) is needed before the “competitive accent reduction” ranking in Table 1 can be treated as fully reliable.
  2. §3.2 and the opening insight of §3: the central controllability claim rests on the assumption that LCS-shared WavLM tokens systematically mark native-like regions. Figs. 4–6 and the progressive ACT/SIM columns of Table 1 supply supporting evidence, yet the paper never reports a direct quantitative check (e.g., phoneme-error or forced-alignment agreement between high-CTP tokens and native targets). Without that check the weakest assumption remains only partially validated; a short diagnostic would strengthen the load-bearing claim that τ is a true accent-strength knob rather than a generic content-preservation knob.
  3. §4.4: the evaluation set contains only 80 sentences. Table 1 and Figs. 5–7 therefore rest on a very small sample; confidence intervals or bootstrap estimates for WER and ΔPPG would make the “lowest WER” ranking more robust, especially given the multi-accent design.
minor comments (4)
  1. Fig. 1 caption and §3: the concatenation of content features with source-token embeddings for CTP/DP is mentioned in text but omitted from the figure; a brief note or dashed arrow would remove ambiguity.
  2. Eq. (12): the loss weights β1=1, β2=1, β3=0.2 are stated only in §4.3; moving them next to the equation would improve readability.
  3. Typographical slips: “arereference-based” (p. 1), “soruce tokens” (Alg. 1 description), “processs” (Fig. 1 caption), “Forth” → “Fourth” (§6).
  4. §5.2.2: the observation that ΔPPG and human ACT diverge under duration preservation is interesting; a one-sentence discussion of why segmental PPG may under-weight rhythm would help readers interpret the metric.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circular derivation; mild evaluation-overlap risk only, not definitional circularity.

full rationale

The paper’s central chain is empirical systems work, not a first-principles derivation that collapses into its inputs. Discrete diffusion follows the external LLaDA ELBO-style masked objective (Eq. 7); CTP is supervised on LCS-derived binary labels between paired source/target tokens (Eq. 8, Fig. 3) and is then used as a control knob, with validity checked by CTP-vs-random ablations (Fig. 5), monotonic τ curves (Fig. 6), and external human ACT/SIM. Duration ratio prediction is standard conditional flow matching on observed ratios (Eq. 10). Content preservation is measured by an external native ASR (WER), naturalness by UTMOSv2 and 25-rater MUSHRA, and speaker similarity by Resemblyzer SECS/BWS—none of which are fitted parameters renamed as predictions. Author-overlapping baselines (TokAN, CosyAccent) are comparison systems, not uniqueness theorems or load-bearing premises that force the result. The only mild concern is that default ΔPPG is computed against the same class of TTS-synthesized native targets used in fine-tuning supervision, which can inflate absolute accent-reduction scores for all systems trained that way; this is evaluation-set affinity, not a self-definitional or fitted-input circular step in the claimed mechanism. No Eq. X reduces to Eq. Y by construction; score remains near zero.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central empirical claim rests on standard discrete-diffusion and flow-matching machinery plus domain assumptions about SSL token phonetics and LCS labeling; free parameters are ordinary training/inference knobs. The main invented pieces are the CTP module and the DLM-AN pipeline that couples it to masked diffusion and duration-ratio prediction.

free parameters (6)
  • joint loss weights β1, β2, β3 = 1.0, 1.0, 0.2
    Set by hand to 1.0, 1.0, 0.2 for DP, CTP, and CTC terms in Eq. (12); affect relative emphasis of duration, common-token, and phoneme objectives.
  • CTP positive class weight = 2
    Hand-chosen weight of 2 for binary cross-entropy label balance during CTP training.
  • CFG strengths w_DLM, w1, w2 = 1.0
    Inference guidance scales for token decoder and synthesizer; set to 1.0 each without reported sweep against the final claim.
  • diffusion / flow sampling steps T = 32
    Greedy unmasking and Euler steps fixed at 32 for reported results.
  • K-means codebook size = 1024
    1024 clusters on WavLM layer-22 features; discrete vocabulary size is a design choice that conditions all token modeling.
  • CTP reuse threshold τ = varied (0.0–1.0)
    User/control hyperparameter (e.g. 1.0, 0.3, 0.0) that directly sets accent strength; not fitted but load-bearing for the controllability claim.
assumptions (5)
  • domain assumption Masked absorbing discrete diffusion with the LLaDA-style reweighted cross-entropy (Eq. 7) is a valid generative model for speech token sequences.
    Imported from LLaDA / discrete diffusion literature (§2.3, §3.1); not re-derived for speech.
  • domain assumption WavLM discrete tokens are sufficiently phonetic that native and L2 renditions share tokens in similarly pronounced regions and differ mainly where accent is strong.
    Stated as the key insight enabling CTP (§3 intro, §3.2); supported by prior SSL analyses but not proved for this tokenizer/accent set.
  • ad hoc to paper Longest common subsequence (with center alignment for runs) between source and target token sequences yields correct binary labels for “common/native-like” tokens.
    Label construction procedure in §3.2 and Fig. 3; LCS was previously used for evaluation, not as CTP supervision.
  • domain assumption TTS-synthesized pseudo-native targets (Matcha-TTS) and source-synthesis L2 counterparts are adequate supervision and evaluation references for accent normalization.
    Data pipeline in §4.1; shared with CosyAccent-style training; quality limits are acknowledged in the introduction but still load-bearing.
  • domain assumption Conditional flow matching correctly models the global duration ratio r = dur_tgt / dur_src from content features.
    §3.3; standard flow-matching assumption applied to a scalar ratio.
invented entities (2)
  • Common Token Predictor (CTP)
    purpose: Sequence tagger that scores source tokens for reuse to control accent strength during reverse diffusion initialization.
    New module trained with LCS-derived labels; no independent external validation beyond this paper’s control curves.
  • DLM-AN pipeline
    purpose: End-to-end controllable accent normalization coupling token encoder, CTP, duration-ratio predictor, masked DLM decoder, and flow-matching synthesizer.
    System-level invention; evidence is internal experimental comparison only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Controllable Accent Normalization via Discrete Diffusion." pith.science (2026). https://pith.science/paper/P3AEW5ED

@misc{pith2026260314275,
  author       = {Pith},
  title        = {Pith review of: Controllable Accent Normalization via Discrete Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3AEW5ED}},
  note         = {Machine review of arXiv:2603.14275}
}
read the original abstract

Existing accent normalization methods do not typically offer control over accent strength, yet many applications-such as language learning and dubbing-require tunable accent retention. We propose DLM-AN, a controllable accent normalization system built on masked discrete diffusion over self-supervised speech tokens. A Common Token Predictor identifies source tokens that likely encode native pronunciation; these tokens are selectively reused to initialize the reverse diffusion process. This provides a simple yet effective mechanism for controlling accent strength: reusing more tokens preserves more of the original accent. DLM-AN further incorporates a flow-matching Duration Ratio Predictor that automatically adjusts the total duration to better match the native rhythm. Experiments on multi-accent English data show that DLM-AN achieves the lowest word error rate among all compared systems while delivering competitive accent reduction and smooth, interpretable accent strength control.

Figures

Figures reproduced from arXiv: 2603.14275 by the authors.

Figure 1
Figure 1. Overview of the DLM-AN pipeline. The SSL tokenizer extracts discrete tokens from L2-accented speech. A Transformer token encoder with CTC-based phonemic guidance produces content representations, which are fed into the Common Token Predictor (CTP), Duration Ratio Predictor (DP), and the DLM decoder. The DLM decoder iteratively generates the target token sequence, optionally initialized with high-CTP-confidence sourc… view at source ↗
Figure 2
Figure 2. Structure of the DLM decoder. The input consists of the content features and the noised target tokens, separated by special tokens [START], [TASK], and [END]. The content features are mutually attentive but do not attend to the token sequence (pink region), while the token sequence attends to the entire input (green region). content features are mutually attentive but with no attention to the token sequence. In cont… view at source ↗
Figure 3
Figure 3. Extraction of common token labels via the longest common subsequence (LCS) between source and target token sequences. For consecutive identical tokens with differing du￾rations, center-mode alignment is applied (dashed rectangle). problem. Given paired source and target token sequences, we derive binary labels by computing the longest common subse￾quence (LCS) between them. LCS has been used previously to evaluate a… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of common token prediction for a Chinese-accented sample. CTP confidence values are overlaid on the Mel￾spectrogram. PPG-predicted phonemes are shown below and their boundaries (white dashed lines) are overlaid on the spectrogram. Aligned words are shown …
Figure 5
Figure 5. Figure 5: CTP-based vs. random token selection at varying reuse proportions. Three metrics are compared: (a) WER, (b) ∆PPG with the L1-accented target, and (c) ∆PPG with the L2-accented source. CTP-based selection achieves generally lower WER and consistently better accent separ…
Figure 6
Figure 6. Figure 6: ∆PPG with the L2-accented source and L1-accented target at varying CTP thresholds τ . Lower τ retains more source tokens. As τ increases, ∆PPG with the source increases (more accent removed) while ∆PPG with the target decreases (closer to native) [PITH_FULL_IMAGE:figu…
Figure 7
Figure 7. Figure 7: shows the results. DLM-AN achieves the low￾est WER (i.e., best content preservation) when the source dura￾tion is preserved, and its advantage is more pronounced when the specified ratio is smaller than 1.0. When the target dura￾tion is set to half of the source, TokAN…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

80 extracted references · 8 linked inside Pith

  1. [1]

    source-synthesis

    Introduction Accent conversion (AC) seeks to alter speech from one accent to another while preserving the speaker’s characteristics. A spe- cial case, accent normalization (AN)2, converts non-native (L2) accented speech into a native (L1) accented form. AN tech- nology enables a wide range of applications, including pronun- ciation training for language l...

  2. [2]

    text-like

    Related Work 2.1. Controllability in Accent Conversion Most accent conversion/normalization systems perform a one- shot “full” accent shift without a user-controllable knob [5, 20]. Recently, controllability has attracted increasing attention, mo- tivated by applications such as gradual accent reduction in lan- guage learning and adjustable accent retenti...

    arXiv 2026

  3. [3]

    The method is shortened “DLM- AN”

    Methodology We propose to use a diffusion language model (DLM) for con- trollable accent normalization. The method is shortened “DLM- AN”. Figure 1 shows the pipeline of DLM-AN. The SSL tok- enizer extracts SSL representations from the L2-accented wave- form and quantize the features into discrete tokens. A Trans- former token encoder takes further proces...

  4. [4]

    a”, unclear “had

    =λ δ zi=[MASK] + (1−λ)δ zi=yi 0 (6) which induces a masked index setMand a visible set ¯M= [L]\M. The corrupted sequencezhas corruption setC=M. The modelp θ(y0 |z,c)is a bidirectional Transformer that predicts the original tokens fromz, conditioned on the content representationscfrom the token encoder. The training objective follows LLaDA: L(θ) =−E t,y0,z...

  5. [5]

    Token-to-Speech Synthesis We use a flow-matching speech synthesizer with a vocoder [54] to generate waveforms

    Nsrc Ntgt + 1 2 ▷source index 7:ifi ⋆ ∈ Ithen 8:z (0) j ←y src i⋆ 9:else 10:z (0) j ←[MASK] 11:end if 12:end for 13:N mask ← |{j|z (0) j =[MASK]}| 14:T eff ← ⌈N mask/K⌉,s 0 ←max(1, T−T eff + 1)▷start step from reuse proportion 15:fors=s 0 toTdo 16:Compute logits at masked positions: ℓcond(· |z (s−1),c)andℓ uncond(· |z (s−1)) 17:Apply CFG:ℓ cfg ←(1 +w DLM)...

  6. [6]

    Datasets The experiments are conducted on English

    Experimental Setup 4.1. Datasets The experiments are conducted on English. Training uses the English subset of Emilia [59] (Emilia-EN) and the LibriTTS-R corpus [60] with synthesized L2-accented counterparts 3 [21]. We also use the L2-ARCTIC corpus [61] together with four American speakers from ARCTIC [62]. We further synthe- size pseudo native targets fo...

  7. [7]

    When provided with the aver- age token duration, TokAN is able to preserve the total du- ration

    the average token duration. When provided with the aver- age token duration, TokAN is able to preserve the total du- ration. We test two modes:T okAN-1, which predicts token durations directly, andT okAN-2, which predicts with total- duration awareness and preserves the total duration. •CosyAccent[21]: A non-autoregressive direct flow-matching model. It f...

    arXiv

  8. [8]

    a” is heav- ily lengthened; correspondingly, the CTP confidence becomes low in the prolonged part. 2) The PPG-predicted phonemes for the word “had

    Results 5.1. Effectiveness of Common Token Prediction For effective common token prediction, higher confidence should be assigned to native-accented regions, whereas low confidence scores should be assigned to highly-L2-accented re- gions. Figure 4 is a visualization of common token predic- tion for a Chinese-accented sample. The common token con- fidence...

Show all 80 references
  1. [9]

    Conclusion & Future Work We presented DLM-AN, a controllable accent normalization system based on masked discrete diffusion over self-supervised speech tokens. By introducing a Common Token Predictor (CTP) that identifies source tokens likely shared with the na- tive target, D...

  2. [10]

    No part of the scientific content, including the ideas, methodology, experiments, or analysis, was generated by AI

    Generative AI Use Disclosure Generative AI tools were used solely for editing and polishing the manuscript text. No part of the scientific content, including the ideas, methodology, experiments, or analysis, was generated by AI. All authors have reviewed and take full responsi...

  3. [11]

    Foreign accent conversion in computer assisted pronunciation training,

    D. Felps, H. Bortfeld, and R. Gutierrez-Osuna, “Foreign accent conversion in computer assisted pronunciation training,”Speech communication, vol. 51, no. 10, pp. 920–932, 2009

    2009

  4. [12]

    Subband based voice conversion

    O. T ¨urk and L. M. Arslan, “Subband based voice conversion.” in Proc. Interspeech, 2002, pp. 289–292

    2002

  5. [13]

    Personalized, cross-lingual tts using phonetic posteriorgrams

    L. Sun, H. Wang, S. Kang, K. Li, and H. M. Meng, “Personalized, cross-lingual tts using phonetic posteriorgrams.” inProc. Inter- speech, 2016, pp. 322–326

    2016

  6. [14]

    Accent conversion using phonetic posteriorgrams,

    Z. Guanlong, S. Sinem, L. John, C.-H. Evgeny, and G.-O. Ri- cardo, “Accent conversion using phonetic posteriorgrams,” in Proc. ICASSP, 2018, pp. 5314–5318

    2018

  7. [15]

    Foreign accent con- version by synthesizing speech from phonetic posteriorgrams

    G. Zhao, S. Ding, and R. Gutierrez-Osuna, “Foreign accent con- version by synthesizing speech from phonetic posteriorgrams.” in Proc. Interspeech, 2019, pp. 2843–2847

    2019

  8. [16]

    Improving accent conversion with ref- erence encoder and end-to-end text-to-speech,

    W. Li, B. Tang, X. Yin, Y . Zhao, W. Li, K. Wang, H. Huang, Y . Wang, and Z. Ma, “Improving accent conversion with ref- erence encoder and end-to-end text-to-speech,”arXiv preprint arXiv:2005.09271, 2020

    2005 arXiv

  9. [17]

    Accentron: Foreign accent conversion to arbitrary non-native speakers using zero-shot learning,

    S. Ding, G. Zhao, and R. Gutierrez-Osuna, “Accentron: Foreign accent conversion to arbitrary non-native speakers using zero-shot learning,”Computer Speech & Language, vol. 72, p. 101302, 2022

    2022

  10. [18]

    Converting foreign accent speech without a reference,

    G. Zhao, S. Ding, and R. Gutierrez-Osuna, “Converting foreign accent speech without a reference,”TASLP, vol. 29, pp. 2367– 2381, 2021

    2021

  11. [19]

    Accent conversion using pre-trained model and synthesized data from voice conver- sion

    T.-N. Nguyen, N.-Q. Pham, and A. Waibel, “Accent conversion using pre-trained model and synthesized data from voice conver- sion.” inProc. Interspeech, 2022, pp. 2583–2587

    2022

  12. [20]

    Zero-shot foreign accent conversion without a native reference,

    W. Quamer, A. Das, J. Levis, E. Chukharev-Hudilainen, and R. Gutierrez-Osuna, “Zero-shot foreign accent conversion without a native reference,” inProc. Interspeech, 2022, pp. 4920–4924

    2022

  13. [21]

    End-to-end accent conversion without using native utterances,

    S. Liu, D. Wang, Y . Cao, L. Sun, X. Wu, S. Kang, Z. Wu, X. Liu, D. Su, D. Yuet al., “End-to-end accent conversion without using native utterances,” inProc. ICASSP, 2020, pp. 6289–6293

    2020

  14. [22]

    V oice-preserving zero-shot multiple accent conversion,

    M. Jin, P. Serai, J. Wu, A. Tjandra, V . Manohar, and Q. He, “V oice-preserving zero-shot multiple accent conversion,” inProc. ICASSP, 2023

    2023

  15. [23]

    Tts-guided train- ing for accent conversion without parallel data,

    Y . Zhou, Z. Wu, M. Zhang, X. Tian, and H. Li, “Tts-guided train- ing for accent conversion without parallel data,”Signal Process- ing Letters, vol. 30, pp. 533–537, 2023

    2023

  16. [24]

    Transfer the linguis- tic representations from tts to accent conversion with non-parallel data,

    X. Chen, J. Pei, L. Xue, and M. Zhang, “Transfer the linguis- tic representations from tts to accent conversion with non-parallel data,” inProc. ICASSP, 2024

    2024

  17. [25]

    Diffusion-based method with tts guidance for foreign accent con- version,

    Q. Bai, S. Wang, Z. Liu, M. Zhang, W. Rao, Y . Wang, and H. Li, “Diffusion-based method with tts guidance for foreign accent con- version,” inProc. ISCSLP, 2024, pp. 284–288

    2024

  18. [26]

    Improving pronunciation and accent conversion through knowledge distilla- tion and synthetic ground-truth from native tts,

    T. N. Nguyen, S. Akti, N. Q. Pham, and A. Waibel, “Improving pronunciation and accent conversion through knowledge distilla- tion and synthetic ground-truth from native tts,” inICASSP, 2025

    2025

  19. [27]

    Accent conversion us- ing discrete units with parallel data synthesized from controllable accented tts,

    T.-N. Nguyen, Q. Pham, and A. Waibel, “Accent conversion us- ing discrete units with parallel data synthesized from controllable accented tts,” inSynthetic Data’s Transformative Role in Founda- tional Speech Models, 2024, pp. 51–55

    2024

  20. [28]

    Convert and speak: Zero- shot accent conversion with minimum supervision,

    Z. Jia, H. Xue, X. Peng, and Y . Lu, “Convert and speak: Zero- shot accent conversion with minimum supervision,” inMultime- dia, 2024

    2024

  21. [29]

    Ac- cent normalization using self-supervised discrete tokens with non- parallel data,

    Q. Bai, S. Inoue, S. Wang, Z. Jiang, Y . Wang, and H. Li, “Ac- cent normalization using self-supervised discrete tokens with non- parallel data,” inInterspeech 2025, 2025, pp. 1618–1622

    2025

  22. [30]

    Vevo: Controllable zero-shot voice imitation with self- supervised disentanglement,

    X. Zhang, X. Zhang, K. Peng, Z. Tang, V . Manohar, Y . Liu, J. Hwang, D. Li, Y . Wang, J. Chan, Y . Huang, Z. Wu, and M. Ma, “Vevo: Controllable zero-shot voice imitation with self- supervised disentanglement,” inICLR, 2025

    2025

  23. [31]

    Cosyac- cent: Duration-controllable accent normalization using source- synthesis training data,

    Q. Bai, S. Shi, S. Wang, Y . Ju, Y . Wang, and H. Li, “Cosyac- cent: Duration-controllable accent normalization using source- synthesis training data,”Proc. ICASSP 2026, 2026

    2026

  24. [32]

    Fac-facodec: Controllable zero-shot foreign accent conversion with factorized speech codec,

    Y . Halychanskyi, C. Churchwell, Y . Wen, and V . Kindratenko, “Fac-facodec: Controllable zero-shot foreign accent conversion with factorized speech codec,”Proc. ICASSP 2026, 2026

    2026

  25. [33]

    Large language diffusion models,

    S. Nie, F. Zhu, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y . Lin, J.-R. Wen, and C. Li, “Large language diffusion models,”arXiv preprint arXiv:2502.09992, 2025

    2025 arXiv

  26. [34]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” inICML, 2006

    2006

  27. [35]

    Non- parallel accent transfer based on fine-grained controllable accent modelling,

    L. Wang, Z. Yu, Y . Yang, S. Gao, C. Mao, and Y . Huang, “Non- parallel accent transfer based on fine-grained controllable accent modelling,” inEMNLP 2023. Association for Computational Linguistics, 2023, pp. 9288–9298

    2023

  28. [36]

    Controllable accented text- to-speech synthesis with fine and coarse-grained intensity render- ing,

    R. Liu, B. Sisman, G. Gao, and H. Li, “Controllable accented text- to-speech synthesis with fine and coarse-grained intensity render- ing,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, p. 2188–2201, Apr. 2024

    2024

  29. [37]

    Scalable control- lable accented tts,

    H. L. Xinyuan, Z. Cai, A. Garg, K. Duh, L. P. Garc ´ıa-Perera, S. Khudanpur, N. Andrews, and M. Wiesner, “Scalable control- lable accented tts,” inProc. ASRU 2025, 2025

    2025

  30. [38]

    Hubert: Self-supervised speech repre- sentation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “Hubert: Self-supervised speech repre- sentation learning by masked prediction of hidden units,”TASLP, vol. 29, pp. 3451–3460, 2021

    2021

  31. [39]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”J-STSP, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  32. [40]

    Self-supervised speech representations are more phonetic than semantic,

    K. Choi, A. Pasad, T. Nakamura, S. Fukayama, K. Livescu, and S. Watanabe, “Self-supervised speech representations are more phonetic than semantic,” inProc. Interspeech, 2024

    2024

  33. [41]

    Any-to-one sequence- to-sequence voice conversion using self-supervised discrete speech representations,

    W.-C. Huang, Y .-C. Wu, and T. Hayashi, “Any-to-one sequence- to-sequence voice conversion using self-supervised discrete speech representations,” inProc. ICASSP, 2021

    2021

  34. [42]

    Textless speech emotion conversion using discrete & decom- posed representations,

    F. Kreuk, A. Polyak, J. Copet, E. Kharitonov, T.-A. Nguyen, M. Rivi `ere, W.-N. Hsu, A. Mohamed, E. Dupoux, and Y . Adi, “Textless speech emotion conversion using discrete & decom- posed representations,” inProc. EMNLP, 2022

    2022

  35. [43]

    Durflex-evc: Duration-flexible emotional voice conversion leveraging discrete representations without text alignment,

    H.-S. Oh, S.-H. Lee, D.-H. Cho, and S.-W. Lee, “Durflex-evc: Duration-flexible emotional voice conversion leveraging discrete representations without text alignment,”IEEE Transactions on Af- fective Computing, 2025

    2025

  36. [44]

    Speak, read and prompt: High-fidelity text-to-speech with mini- mal supervision,

    E. Kharitonov, D. Vincent, Z. Borsos, R. Marinier, S. Girgin, O. Pietquin, M. Sharifi, M. Tagliasacchi, and N. Zeghidour, “Speak, read and prompt: High-fidelity text-to-speech with mini- mal supervision,”Trans. ACL, vol. 11, pp. 1703–1718, 2023

    2023

  37. [45]

    On generative spoken language modeling from raw audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, Y . Adi, A. Polyak, B. Bolte, T.-A. Nguyen, J. Copet, A. Baevski, A. Mohamed et al., “On generative spoken language modeling from raw audio,” Trans. ACL, vol. 9, pp. 1336–1354, 2021

    2021

  38. [46]

    Direct speech- to-speech translation with discrete units,

    A. Lee, P.-J. Chen, C. Wang, J. Gu, S. Popuri, X. Ma, A. Polyak, Y . Adi, Q. He, Y . Tang, J. Pino, and W.-N. Hsu, “Direct speech- to-speech translation with discrete units,” inProc. ACL, 2022

    2022

  39. [47]

    Llama- omni: Seamless speech interaction with large language models,

    Q. Fang, S. Guo, Y . Zhou, Z. Ma, S. Zhang, and Y . Feng, “Llama- omni: Seamless speech interaction with large language models,” arXiv preprint arXiv:2409.06666, 2024

    2024 arXiv

  40. [48]

    Exploring ssl discrete tokens for multilingual asr,

    M. Cui, D. Tan, Y . Yang, D. Wang, H. Wang, X. Chen, X. Chen, and X. Liu, “Exploring ssl discrete tokens for multilingual asr,” in Proc. ICASSP 2025, 2025

    2025

  41. [49]

    Exploring ssl discrete speech features for zipformer-based contextual asr,

    M. Cui, Y . Yang, J. Deng, J. Kang, S. Hu, T. Wang, Z. Li, S. Zhang, X. Chen, and X. Liu, “Exploring ssl discrete speech features for zipformer-based contextual asr,” inProc. Interspeech 2025, 2025

    2025

  42. [50]

    Advanced modeling of interlanguage speech intelligibility benefit with l1-l2 multi-task learning using differentiable k-means for accent-robust discrete token-based asr,

    K. Onda, S. Fukayama, D. Saito, and N. Minematsu, “Advanced modeling of interlanguage speech intelligibility benefit with l1-l2 multi-task learning using differentiable k-means for accent-robust discrete token-based asr,” inProc. ICASSP 2026, 2026

    2026

  43. [51]

    Argmax flows and multinomial diffusion: Learning categorical distributions,

    E. Hoogeboom, D. Nielsen, P. Jaini, P. Forr ´e, and M. Welling, “Argmax flows and multinomial diffusion: Learning categorical distributions,” inAdvances in neural information processing sys- tems, vol. 34, 2021, pp. 12 454–12 465

    2021

  44. [52]

    Structured denoising diffusion models in discrete state-spaces,

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. van den Berg, “Structured denoising diffusion models in discrete state-spaces,” inAdvances in Neural Information Processing Systems, 2021

    2021

  45. [53]

    Maskgit: Masked generative image transformer,

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman, “Maskgit: Masked generative image transformer,” inProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 315–11 325

    2022

  46. [54]

    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,” inThe Thirteenth International Confer- ence on Learning Representations, 2025. [Online]. Available: http...

    2025

  47. [55]

    Metis: A foundation speech generation model with masked generative pre-training,

    Y . Wang, J. Zheng, J. Zhang, X. Zhang, H. Liao, and Z. Wu, “Metis: A foundation speech generation model with masked generative pre-training,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum...

    2025

  48. [56]

    Discrete diffusion in large language and multimodal models: A survey,

    R. Yu, Q. Li, and X. Wang, “Discrete diffusion in large language and multimodal models: A survey,” 2025. [Online]. Available: https://arxiv.org/abs/2506.13759

    2025

  49. [57]

    Discrete diffusion modeling by estimating the ratios of the data distribution,

    A. Lou, C. Meng, and S. Ermon, “Discrete diffusion modeling by estimating the ratios of the data distribution,” inICML 2024, 2024

    2024

  50. [58]

    Re- masking discrete diffusion models with inference-time scal- ing,

    G. Wang, Y . Schiff, S. S. Sahoo, and V . Kuleshov, “Re- masking discrete diffusion models with inference-time scal- ing,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id=IJryQAOy0p

    2025

  51. [59]

    Accelerated sampling from masked diffusion models via entropy bounded unmasking,

    H. Ben-Hamu, I. Gat, D. Severo, N. Nolte, and B. Karrer, “Accelerated sampling from masked diffusion models via entropy bounded unmasking,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id...

    2025

  52. [60]

    Don’t settle too early: Self-reflective remasking for diffusion language models,

    Z. Huang, Y . Wang, Z. Chen, and G.-J. Qi, “Don’t settle too early: Self-reflective remasking for diffusion language models,”arXiv preprint arXiv:2509.23653, 2025

    2025

  53. [61]

    Corrective diffusion language models,

    S. Zhang, F. Z. Peng, Y . Zhang, J. Pan, and G. G. Chrysos, “Corrective diffusion language models,” 2026. [Online]. Available: https://arxiv.org/abs/2512.15596

    2026

  54. [62]

    Seed diffusion: A large-scale dif- fusion language model with high-speed inference,

    Y . Song, Z. Zhang, C. Luo, P. Gao, F. Xia, H. Luo, Z. Li, Y . Yang, H. Yu, X. Quet al., “Seed diffusion: A large-scale dif- fusion language model with high-speed inference,”arXiv preprint arXiv:2508.02193, 2025

    2025 arXiv

  55. [63]

    Llada2. 1: Speeding up text diffusion via token editing,

    T. Bie, M. Cao, X. Cao, B. Chen, F. Chen, K. Chen, L. Du, D. Feng, H. Feng, M. Gonget al., “Llada2. 1: Speeding up text diffusion via token editing,”arXiv preprint arXiv:2602.08676, 2026

    2026

  56. [64]

    Hiftnet: A fast high-quality neural vocoder with harmonic-plus-noise fil- ter and inverse short time fourier transform,

    Y . A. Li, C. Han, X. Jiang, and N. Mesgarani, “Hiftnet: A fast high-quality neural vocoder with harmonic-plus-noise fil- ter and inverse short time fourier transform,”arXiv preprint arXiv:2309.09493, 2023

    2023 arXiv

  57. [65]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in NeurIPS 2021 Workshop on Deep Generative Models and Down- stream Applications, 2021

    2021

  58. [66]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inProc. ICCV, 2023, pp. 4195–4205

    2023

  59. [67]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inICLR, 2023

    2023

  60. [68]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: Enhanced transformer with rotary position embedding,”Neuro- computing, vol. 568, p. 127063, 2024

    2024

  61. [69]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shiet al., “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 885–890

    2024

  62. [70]

    LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, M. Bacchiani, Y . Zhang, W. Han, and A. Bapna, “LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus,” inProc. Inter- speech, 2023

    2023

  63. [71]

    L2-ARCTIC: A Non-native English Speech Corpus,

    G. Zhao, S. Sonsaat, A. Silpachai, I. Lucic, E. Chukharev- Hudilainen, J. Levis, and R. Gutierrez-Osuna, “L2-ARCTIC: A Non-native English Speech Corpus,” inProc. Interspeech, 2018

    2018

  64. [72]

    The cmu arctic speech databases,

    J. Kominek and A. W. Black, “The cmu arctic speech databases,” inFifth ISCA workshop on speech synthesis, 2004

    2004

  65. [73]

    Matcha-tts: A fast tts architecture with conditional flow match- ing,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-tts: A fast tts architecture with conditional flow match- ing,” inProc. ICASSP, 2024, pp. 11 341–11 345

    2024

  66. [74]

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension,

    M. Lewis, “Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension,” arXiv preprint arXiv:1910.13461, 2019

    1910 arXiv

  67. [75]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inICML, 2023, pp. 28 492–28 518

    2023

  68. [76]

    Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    2024 arXiv

  69. [77]

    A comparison of best-worst scaling and rat- ing scale for timbre characterisation,

    A. M. V . Ravillion, “A comparison of best-worst scaling and rat- ing scale for timbre characterisation,” 2020

    2020

  70. [78]

    High-fidelity neural phonetic posteriorgrams,

    C. Churchwell, M. Morrison, and B. Pardo, “High-fidelity neural phonetic posteriorgrams,” inICASSP 2024 Workshop on Explain- able Machine Learning for Speech and Audio, 2024

    2024

  71. [79]

    Montreal forced aligner: Trainable text-speech align- ment using kaldi

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal forced aligner: Trainable text-speech align- ment using kaldi.” inProc. Interspeech, 2017, pp. 498–502

    2017

  72. [80]

    Neural dis- crete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural dis- crete representation learning,” inProceedings of the 31st Inter- national Conference on Neural Information Processing Systems, 2017, p. 6309–6318

    2017

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.