Pith. sign in

REVIEW 2 major objections 5 minor 25 references

A lightweight adaptation turns a frozen autoregressive TTS model into a confidence-ordered diffusion decoder that is both more accurate and 3.3 imes faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 21:23 UTC pith:CUBBUHJA

load-bearing objection Solid first transfer of AR-to-dLLM conversion to discrete speech tokens: real WER and speed gains on limited data, with the main soft spot already quantified by the authors. the 2 major comments →

arxiv 2607.04140 v1 pith:CUBBUHJA submitted 2026-07-05 eess.AS cs.CL

DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech

Junwon Moon , Seungbeom Kim , Yejin Lee , Hoseong Ahn , Sewoong Park , Heeseung Kim , Kyuhong Shim This is my paper
classification eess.AS cs.CL
keywords text-to-speechdiscrete diffusion language modelautoregressive-to-diffusion conversionconfidence-ordered decodingLoRA adaptationspeech token generationnon-autoregressive TTS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Autoregressive text-to-speech models must emit every speech token left-to-right, so each decision is made without seeing future acoustic context; that fixed order slows inference and lets early mistakes cascade into hallucinations. The paper shows that this ordering is not required: because the full text is already known, a pretrained AR backbone can be converted, with only about 15 % extra parameters, into a discrete diffusion language model that unmasks tokens in confidence order. A small convolution module supplies the local acoustic mixing that pure attention lacks, while a 1/t training weight and a time-shifted schedule deliberately postpone uncertain positions until surrounding tokens have been resolved. Trained on just 585 hours of public English speech, the resulting system reaches 1.75 % word-error rate—better than its own AR teacher—while generating tokens 3.3 times faster and largely eliminating the catastrophic early-commitment failures that still afflict the original model.

Core claim

A frozen autoregressive TTS backbone can be turned into a non-autoregressive discrete diffusion model by bidirectional attention plus lightweight LoRA and convolution adapters; the resulting confidence-ordered decoder produces sharper text–speech alignment, higher average token confidence, and lower word-error rate than the original left-to-right model while running several times faster.

What carries the argument

DELTA-TTS: a LoRA-plus-Conformer-convolution adapter stack that re-uses the AR next-token head under a masked-diffusion objective, combined with a 1/t-weighted loss and a time-shifted unmasking schedule that commits high-confidence positions first.

Load-bearing premise

The length of the fully masked speech-token sequence can be set accurately enough by a simple rule that scales the prompt’s audio-to-character rate by the target character count.

What would settle it

Replace the rule-based length estimator with an oracle that supplies the true ground-truth token count (or a deliberately wrong length) and measure whether the reported WER advantage and hallucination reduction disappear.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any AR TTS model whose weights are already strong can be converted into a parallel decoder with only modest extra training data and parameters.
  • Confidence-ordered unmasking systematically defers the low-confidence positions that cause AR hallucination cascades, giving a structural robustness gain.
  • Local convolution is necessary to restore the short-range acoustic structure that pure bidirectional attention fails to capture from AR weights alone.
  • The same conversion recipe should transfer to other AR speech language models and multi-codebook tokenizers once length prediction is improved.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adapter pattern may let other left-to-right generative models (music, video tokens) abandon fixed causal order without full retraining.
  • If length estimation remains the main remaining bottleneck, discrete diffusion TTS will stay English-centric until a learned duration head is added.
  • Reward fine-tuning that directly optimizes WER and speaker similarity could further close the remaining gap to the strongest pure AR systems.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. DELTA-TTS converts a frozen CosyVoice3 AR TTS backbone into a discrete diffusion language model via bidirectional attention, block-wise LoRA, and a Conformer-style convolution residual, trained with a 1/t-weighted masked diffusion objective and decoded with a time-shifted confidence-ordered unmasking schedule. On only 585 h of LibriTTS the adapted model reports 1.75 % WER on Seed-TTS test-en (vs 2.02 % for the AR backbone) at 3.3 imes token-generation speedup, competitive SIM/UTMOS, and subjective gains; ablations and attention/confidence analyses attribute the gains to local mixing and deferred low-confidence commitments that reduce AR-style hallucinations.

Significance. If the empirical claims hold, the work supplies a practical, parameter-efficient route (≈15 % extra parameters) for turning existing LLM-backed AR TTS systems into non-autoregressive generators without full retraining or large-scale speech corpora. The speech-specific design choices (convolution for locality, 1/t + time-shift for confidence ordering) and the supporting analyses (sharper text–speech alignment, elevated commit confidence, concrete hallucination cases) make a concrete case that confidence-ordered decoding is more robust than left-to-right commitment for speech tokens. The public-data, low-resource setting and clear Pareto comparison against strong AR/NAR baselines further increase the result’s utility for the community.

major comments (2)
  1. Section 4.4 and Appendix A.7: target length is set by a rule-based prompt audio-to-character rate rather than a learned duration model. While the GT-length ablation in Table 5 shows nearly identical WER (1.63 vs 1.75), SIM is substantially lower without prompt conditioning, and the authors themselves flag length estimation as an open dLLM limitation. A short quantitative stress test (WER/SIM under deliberately mismatched lengths, or comparison to a simple duration predictor) would make clear how much of the reported zero-shot quality depends on this heuristic remaining accurate on the evaluation sets.
  2. Tables 1 and 4: the strongest AR baselines (VoxCPM, FireRedTTS2, IndexTTS2, etc.) are trained on 55 k–1800 k hours of multi-domain data, while DELTA-TTS uses 585 h of LibriTTS. The paper correctly notes the data disparity, yet the headline claim of “outperforming its AR backbone” is sometimes read as a general superiority claim. Explicitly framing the result as “surpassing the frozen backbone under a 585 h adaptation budget” and, if possible, reporting a same-data AR fine-tune control would tighten the comparison.
minor comments (5)
  1. Figure 1 caption and surrounding text refer to “palette options / Final palette for the paper”; residual drafting notes should be removed.
  2. Equation (2) and the subsequent inference schedule both use t / µ; a brief sentence clarifying that the continuous-time formulation is discretized only at inference would help readers less familiar with masked diffusion.
  3. Table 3 length buckets are useful; adding the corresponding absolute token counts or average utterance lengths would make the RTF scaling easier to interpret.
  4. Appendix A.6 subjective protocol is transparent; stating the exact number of ratings per system (or total listener-hours) would complete the reproducibility information.
  5. A few typographical inconsistencies remain (e.g., “CosyV oice” spacing, “starget” vs “s_target”).

Circularity Check

0 steps flagged

No significant circularity: empirical AR-to-dLLM adaptation evaluated on external ASR/SIM metrics with no self-referential derivation.

full rationale

DELTA-TTS is an empirical systems paper. The central claims (1.75% WER on Seed-TTS test-en outperforming the frozen CosyVoice3 AR backbone, 3.3 imes token-generation speedup, sharper alignment, higher commit confidence, fewer hallucinations) are measured with independent external tools (Whisper-Large-V3 WER, WavLM-ECAPA SIM, UTMOS, human CMOS/SMOS, A100 RTF). The training objective (Eq. 2) is the standard masked-diffusion ELBO with 1/t weighting taken from the cited literature (Sahoo et al., Nie et al., Gong et al.); the time-shifted schedule (Eq. 3) is an adaptation of Esser et al.; neither quantity is fitted to the reported WER/SIM numbers. The rule-based length estimator is ablated against ground-truth length (Table 5) and does not force the intelligibility gain. Self-citations are limited to the CosyVoice3 backbone used as a frozen starting point; no uniqueness theorem, ansatz, or load-bearing premise is imported from overlapping authors. No equation reduces a claimed prediction to an input by construction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard discrete-diffusion and LoRA machinery plus a handful of speech-specific modeling choices and free hyper-parameters. No new physical entities are postulated; the free parameters are ordinary ML knobs whose values are reported.

free parameters (5)
  • time-shift parameter µ = 0.3
    Controls the cumulative unmasking schedule (Eq. 3); set to 0.3 in experiments; chosen by hand to defer low-confidence tokens.
  • convolution kernel size = 31
    Depth-wise kernel width in the Conformer-style module; fixed at 31; affects local acoustic context.
  • LoRA rank / α = α=128
    Low-rank adapter dimensions and scaling; α=128 stated; rank not explicitly listed but contributes to the 35 M LoRA parameters.
  • number of unmasking steps T = 16
    Inference budget; default 16 after Pareto sweep over {4,8,16}; directly trades quality for speed.
  • rule-based length ratio
    Target speech-token count = r_prompt × W_target; heuristic that replaces a learned duration model.
axioms (4)
  • domain assumption Masked discrete diffusion ELBO with 1/t reweighting is a valid training objective for recovering speech tokens.
    Imported from text dLLM literature (Sahoo et al., Nie et al., Gong et al.) and applied unchanged to 25 Hz speech tokens (Section 3.4).
  • domain assumption Neighboring 25 Hz speech tokens exhibit strong local acoustic correlation that bidirectional attention alone under-models.
    Justifies insertion of the Conformer-style convolution residual (Section 3.3).
  • ad hoc to paper A frozen AR backbone plus low-rank adapters can absorb the switch from causal to bidirectional attention without mask annealing.
    Contrasts with prior text AR-to-dLLM work that required gradual mask annealing; claimed to hold for the CosyVoice3 backbone (Section 3.3).
  • domain assumption Standard next-token shift operation preserves AR pretrained weights under masked prediction.
    Taken from Gong et al. / Ye et al. and reused (Section 3.4).
invented entities (1)
  • DELTA-TTS adapter stack (block-wise LoRA + Conformer convolution) independent evidence
    purpose: Enables AR-to-dLLM conversion for speech while injecting local acoustic bias.
    Composite engineering construct; not a fundamental new particle or force, but the paper’s named technical contribution. Independent evidence is the reported empirical gains on public benchmarks.

pith-pipeline@v1.1.0-grok45 · 19834 in / 3010 out tokens · 26863 ms · 2026-07-11T21:23:59.032075+00:00 · methodology

0 comments
read the original abstract

Autoregressive (AR) text-to-speech (TTS) models generate discrete speech tokens sequentially, which makes inference slow and can degrade robustness by propagating local errors and hallucinations. This limitation stems from their left-to-right AR commitment: each token must be determined before future speech-token context is available. However, such ordering is not an inherent requirement for TTS, as the full input text is available before synthesis. In this paper, we introduce DELTA-TTS, a lightweight LoRA-based adaptation framework that converts a pretrained AR TTS model into a discrete diffusion language model (dLLM) for confidence-ordered speech-token decoding. To better capture the local structure of speech, DELTA-TTS incorporates a convolution module that injects local acoustic context, together with a $1/t$-weighted training objective and a time-shifted inference schedule that defer low-confidence positions to later steps. Trained on only $585$ hours of LibriTTS, DELTA-TTS achieves a $\textbf{1.75}\%$ WER on Seed-TTS test-en, outperforming its AR backbone while generating tokens $\textbf{3.3}\times$ faster. Further analysis shows that DELTA-TTS produces sharper text--speech alignment, increases overall decoding confidence, and mitigates hallucinations observed in AR generation.

Figures

Figures reproduced from arXiv: 2607.04140 by Heeseung Kim, Hoseong Ahn, Junwon Moon, Kyuhong Shim, Seungbeom Kim, Sewoong Park, Yejin Lee.

Figure 1
Figure 1. Figure 1: Overview of DELTA-TTS. A pretrained AR TTS backbone is converted into a discrete diffusion language model with two adaptations: bidirectional attention and block-wise LoRA with a Conformer-style convolution module. We validate DELTA-TTS through both empirical evaluation and analysis. First, naively applying an AR-to-dLLM conver￾sion designed for text degrades performance substantially, whereas our speech-a… view at source ↗
Figure 2
Figure 2. Figure 2: Speed–quality Pareto front on Seed-TTS test-en (both axes log-scale). DELTA-TTS at 16 steps reaches the bottom-left corner, achieving the lowest WER and RTF among all NAR baselines and the AR baseline. The result shows that DELTA-TTS lies on the Pareto front, offering the most favorable speed–quality trade-off among the evaluated systems. formation across the full sequence but has limited inductive bias to… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Text–speech alignment across depth: DELTA-TTS sharpens the middle-layer diagonal and, in the late layer, attends to upcoming text, a pattern unavailable to the causal AR baseline. (b) Per-step attention budget from masked tokens; the convolution module shifts more attention onto neighboring unmasked speech while keeping text attention stable. ring uncertain commitments to later steps. The block-wise co… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Mean teacher-forced AR confidence on the GT speech token across 1088 Seed-TTS test-en utterances; the first 5 positions sit close to zero. (b) Marginal distribution of per-position confidence at the GT speech token; the AR model’s teacher-forced confidence concentrates near zero, while DELTA-TTS commits at far higher confidence [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Number of tokens unmasked per step under uniform and shifted schedules (left), and the 1/t loss weight as a function of the masking ratio (right). A.4. Easy-first, Hard-last Unmasking Order [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-step informative-token (vowel + consonant) ratio averaged over 1088 utterances. Strip on top: phoneme classes on one example utterance. A.5. Per-Step Commit Ordering [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: shows, at each unmasking step, the mean teacher-forced AR confidence of the positions DELTA-TTS commits. DELTA-TTS commits high-AR-confidence positions early and defers low-AR-confidence positions to later steps, once the bidirectional decoder has resolved more surrounding context. This deferral is what lets DELTA-TTS commit even AR-hard positions at high confidence, producing the elevated confidence distr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 22 linked inside Pith

  1. [1]

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

    Anastassiou, P., Chen, J., Chen, J., Chen, Y ., Chen, Z., Chen, Z., Cong, J., Deng, L., Ding, C., Gao, L., et al. Seed-tts: A family of high-quality versatile speech generation models. arXiv preprint arXiv:2406.02430,

    Pith/arXiv arXiv

  2. [2]

    T., Yang, Z., Qi, Z., Han, J., Sahoo, S

    Arriola, M., Gokaslan, A., Chiu, J. T., Yang, Z., Qi, Z., Han, J., Sahoo, S. S., and Kuleshov, V . Block diffusion: Inter- polating between autoregressive and diffusion language models.arXiv preprint arXiv:2503.09573,

    Pith/arXiv arXiv

  3. [3]

    Soundstorm: Efficient par- allel audio generation.arXiv preprint arXiv:2305.09636,

    Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghi- dour, N., and Tagliasacchi, M. Soundstorm: Efficient par- allel audio generation.arXiv preprint arXiv:2305.09636,

    Pith/arXiv arXiv

  4. [4]

    Ecapa- tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification.arXiv preprint arXiv:2005.07143,

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

    Pith/arXiv arXiv 2005

  5. [5]

    H., Doucet, A., Strudel, R., Dyer, C., Durkan, C., et al

    Dieleman, S., Sartran, L., Roshannai, A., Savinov, N., Ganin, Y ., Richemond, P. H., Doucet, A., Strudel, R., Dyer, C., Durkan, C., et al. Continuous diffusion for categorical data.arXiv preprint arXiv:2211.15089,

    Pith/arXiv arXiv

  6. [6]

    Cosyvoice: A scalable multilingual zero-shot text-to-speech synthe- sizer based on supervised semantic tokens.arXiv preprint arXiv:2407.05407, 2024a

    Du, Z., Chen, Q., Zhang, S., Hu, K., Lu, H., Yang, Y ., Hu, H., Zheng, S., Gu, Y ., Ma, Z., et al. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthe- sizer based on supervised semantic tokens.arXiv preprint arXiv:2407.05407, 2024a. Du, Z., Wang, Y ., Chen, Q., Shi, X., Lv, X., Zhao, T., Gao, Z., Yang, Y ., Gao, C., Wang, H., et al. Cosyvo...

    Pith/arXiv arXiv

  7. [7]

    Diffuseq: Sequence to sequence text generation with diffusion mod- els.arXiv preprint arXiv:2210.08933,

    Gong, S., Li, M., Feng, J., Wu, Z., and Kong, L. Diffuseq: Sequence to sequence text generation with diffusion mod- els.arXiv preprint arXiv:2210.08933,

    Pith/arXiv arXiv

  8. [8]

    Scaling diffu- sion language models via adaptation from autoregressive models.arXiv preprint arXiv:2410.17891,

    Gong, S., Agarwal, S., Zhang, Y ., Ye, J., Zheng, L., Li, M., An, C., Zhao, P., Bi, W., Han, J., et al. Scaling diffu- sion language models via adaptation from autoregressive models.arXiv preprint arXiv:2410.17891,

    Pith/arXiv arXiv

  9. [9]

    Con- former: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100,

    Gulati, A., Qin, J., Chiu, C.-C., Parmar, N., Zhang, Y ., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y ., et al. Con- former: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100,

    Pith/arXiv arXiv 2005

  10. [10]

    Fireredtts-1s: An upgraded stream- able foundation text-to-speech system.arXiv preprint arXiv:2503.20499,

    Guo, H.-H., Hu, Y ., Shen, F.-Y ., Tang, X., Wu, Y .-C., Xie, F.-L., and Xie, K. Fireredtts-1s: An upgraded stream- able foundation text-to-speech system.arXiv preprint arXiv:2503.20499,

    Pith/arXiv arXiv

  11. [11]

    The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751,

    Holtzman, A., Buys, J., Du, L., Forbes, M., and Choi, Y . The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751,

    Pith/arXiv arXiv 1904

  12. [12]

    Qwen3-tts technical report.arXiv preprint arXiv:2601.15621,

    Hu, H., Zhu, X., He, T., Guo, D., Zhang, B., Wang, X., Guo, Z., Jiang, Z., Hao, H., Guo, Z., et al. Qwen3-tts technical report.arXiv preprint arXiv:2601.15621,

    Pith/arXiv arXiv

  13. [13]

    Ditar: Diffusion trans- former autoregressive modeling for speech generation

    Jia, D., Chen, Z., Chen, J., Du, C., Wu, J., Cong, J., Zhuang, X., Li, C., Wei, Z., Wang, Y ., et al. Ditar: Diffusion trans- former autoregressive modeling for speech generation. arXiv preprint arXiv:2502.03930,

    arXiv

  14. [14]

    Naturalspeech 3: Zero- shot speech synthesis with factorized codec and diffusion models.arXiv preprint arXiv:2403.03100,

    9 DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for TTS Ju, Z., Wang, Y ., Shen, K., Tan, X., Xin, D., Yang, D., Liu, Y ., Leng, Y ., Song, K., Tang, S., et al. Naturalspeech 3: Zero- shot speech synthesis with factorized codec and diffusion models.arXiv preprint arXiv:2403.03100,

    Pith/arXiv arXiv

  15. [15]

    W., Kim, J., Chung, S., and Cho, J

    Lee, K., Kim, D. W., Kim, J., Chung, S., and Cho, J. Ditto-tts: Diffusion transformers for scalable text-to- speech without domain-specific factors.arXiv preprint arXiv:2406.11427,

    Pith/arXiv arXiv

  16. [16]

    Discrete diffusion model- ing by estimating the ratios of the data distribution.arXiv preprint arXiv:2310.16834,

    Lou, A., Meng, C., and Ermon, S. Discrete diffusion model- ing by estimating the ratios of the data distribution.arXiv preprint arXiv:2310.16834,

    Pith/arXiv arXiv

  17. [17]

    Large language diffusion models.arXiv preprint arXiv:2502.09992,

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

    Pith/arXiv arXiv

  18. [18]

    Utmos: Utokyo-sarulab sys- tem for voicemos challenge 2022.arXiv preprint arXiv:2204.02152,

    Saeki, T., Xin, D., Nakata, W., Koriyama, T., Takamichi, S., and Saruwatari, H. Utmos: Utokyo-sarulab sys- tem for voicemos challenge 2022.arXiv preprint arXiv:2204.02152,

    Pith/arXiv arXiv 2022

  19. [19]

    Neural codec language models are zero-shot text to speech synthesizers

    Wang, C., Chen, S., Wu, Y ., Zhang, Z., Zhou, L., Liu, S., Chen, Z., Liu, Y ., Wang, H., Li, J., et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111,

    Pith/arXiv arXiv

  20. [20]

    Spark- tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens.arXiv preprint arXiv:2503.01710, 2025a

    Wang, X., Jiang, M., Ma, Z., Zhang, Z., Liu, S., Li, L., Liang, Z., Zheng, Q., Wang, R., Feng, X., et al. Spark- tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens.arXiv preprint arXiv:2503.01710, 2025a. Wang, Y ., Zhan, H., Liu, L., Zeng, R., Guo, H., Zheng, J., Zhang, Q., Zhang, X., Zhang, S., and Wu, Z. Maskgct:...

    Pith/arXiv arXiv

  21. [21]

    Metis: A foundation speech generation model with masked generative pre-training.arXiv preprint arXiv:2502.03128, 2025b

    Wang, Y ., Zheng, J., Zhang, J., Zhang, X., Liao, H., and Wu, Z. Metis: A foundation speech generation model with masked generative pre-training.arXiv preprint arXiv:2502.03128, 2025b. Wu, C., Zhang, H., Xue, S., Diao, S., Fu, Y ., Liu, Z., Molchanov, P., Luo, P., Han, S., and Xie, E. Fast- dllm v2: Efficient block-diffusion llm.arXiv preprint arXiv:2509.26328,

    Pith/arXiv arXiv

  22. [22]

    Fireredtts-2: Towards long conversational speech generation for podcast and chatbot.arXiv preprint arXiv:2509.02020,

    Xie, K., Shen, F., Li, J., Xie, F., Tang, X., and Hu, Y . Fireredtts-2: Towards long conversational speech generation for podcast and chatbot.arXiv preprint arXiv:2509.02020,

    Pith/arXiv arXiv

  23. [23]

    Dream 7b: Diffusion large language models

    Ye, J., Xie, Z., Zheng, L., Gao, J., Wu, Z., Jiang, X., Li, Z., and Kong, L. Dream 7b: Diffusion large language models. arXiv preprint arXiv:2508.15487, 2025a. Ye, Z., Zhu, X., Chan, C.-M., Wang, X., Tan, X., Lei, J., Peng, Y ., Liu, H., Jin, Y ., Dai, Z., et al. Llasa: Scal- ing train-time and inference-time compute for llama- based speech synthesis.arXi...

    Pith/arXiv arXiv 1904

  24. [24]

    Indextts2: A breakthrough in emotionally expres- sive and duration-controlled auto-regressive zero-shot text-to-speech

    Zhou, S., Zhou, Y ., He, Y ., Zhou, X., Wang, J., Deng, W., and Shu, J. Indextts2: A breakthrough in emotionally expres- sive and duration-controlled auto-regressive zero-shot text-to-speech. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pp. 35139–35148, 2026a. Zhou, Y ., Zeng, G., Liu, X., Li, X., Yu, R., Wang, Z., Ye, R., S...

    arXiv

  25. [25]

    dllm: Simple diffusion language modeling.arXiv preprint arXiv:2602.22661, 2026b

    Zhou, Z., Chen, L., Tong, H., and Song, D. dllm: Simple diffusion language modeling.arXiv preprint arXiv:2602.22661, 2026b. 10 DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for TTS A. Appendix A.1. Implementation Details We train on a single NVIDIA A100 with bfloat16 mixed precision. The LoRA targets the query, key, value, output,...

    arXiv