Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

AudioTurbo: Fast Text-to-Audio Generation with Rectified Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read AudioTurbo claims that fine-tuning a pretrained text-to-audio diffusion model on deterministic noise-sample pairs lets it generate higher-quality audio in 10 sampling steps than prior models achieve in 200 steps, and in 3 steps it matches…

desk verdict A useful and honest application of rectified diffusion to text-to-audio, but the abstract overclaims against the teacher model and the guidance scale is tuned on the test set. read the letter →

arxiv 2505.22106 v1 pith:PCCIB7KF submitted 2025-05-28 cs.SD cs.AIeess.AS

Junqi Zhao , Jinzheng Zhao , Haohe Liu , Yun Chen , Lu Han , Xubo Liu , Mark Plumbley , Wenwu Wang This is my paper
classification cs.SDcs.AIeess.AS
keywords text-to-audiogenerationrectifieddiffusionflowmatchinglatentmodelinferenceaccelerationclassifier-freeguidanceAudioCaps
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

AudioTurbo claims that rectified diffusion, a retraining procedure that pairs each noise vector with a fixed generated sample, can be grafted onto an existing pretrained text-to-audio model to cut inference cost dramatically without losing quality. The paper reports that on AudioCaps, AudioTurbo with 10 sampling steps outperforms strong baselines (Tango, AudioLDM2, Auffusion, LAFMA) that use 200 steps, and with 3 steps it matches the best 25-step results of LAFMA, a flow-matching acceleration model. If true, the practical consequence is that high-quality text-to-audio can run at interactive speeds while reusing the investment already made in large pretrained diffusion models. The central move is to replace randomly paired noise and data with pairs produced by a frozen teacher model, so the learned ODE trajectories become consistent and few-step solvers work.

What carries the argument

The load-bearing mechanism is the deterministic pair of noise and generated latent produced by the pretrained Auffusion model, combined with the rectified-diffusion retraining procedure. Rectified diffusion states that if $x_t = \alpha_t x_0 + \sigma_t \epsilon$ with a fixed pair $(\epsilon, x_0)$, the first-order ODE solve of Equation 5 is exact because the prediction $\epsilon_\theta$ is constant along the trajectory. AudioTurbo fine-tunes Auffusion's UNet on 47,744 such pairs, with 10% text dropping for classifier-free guidance at scale 1.5.

What would settle it

Train AudioTurbo with the same recipe but generate the 47,744 training pairs using a different teacher model (for example AudioLDM2 or Make-an-Audio), then check whether the 10-step results still beat 200-step baselines on AudioCaps; if the speed-quality advantage shrinks or reverses, the claimed improvement is tied to Auffusion's specific output distribution rather than to rectified diffusion itself.

Watch

Extended reading notes

Core claim

The paper's central claim is that a pretrained diffusion-based text-to-audio model can be converted into a fast sampler by fine-tuning it under the rectified-diffusion objective: use the frozen teacher model Auffusion to generate audio latents from fixed Gaussian noise, pair each noise with its generated latent, and retrain the UNet on these deterministic pairs. Because the noise and data are now a deterministic coupling, the model's noise prediction stays roughly constant along a trajectory, so a first-order ODE solve with 5 or 10 steps produces audio that is as good as or better than 200-step sampling from the original model. On the AudioCaps test set the paper reports FD 20.65, KL 1.29, IS 9.40, and CLAP 29.8 at 10 steps, all best among the compared models, with subjective ratings (OVL 82.25, REL 85.58) also above baselines.

Load-bearing premise

The entire training set is audio that Auffusion generated from AudioCaps captions, so the method is only as good as the teacher's coverage and faithfulness; if those synthetic samples are narrow or artifact-laden, the reported metrics mostly measure the teacher's biases rather than a genuine text-to-audio ability.

Editorial extensions

If this is right

  • Text-to-audio generation could run at roughly 20 to 40 times fewer sampling steps than current diffusion baselines, making real-time or near-real-time use feasible on consumer hardware.
  • The acceleration does not require training from scratch: only the UNet is fine-tuned while the text encoder, VAE, and vocoder stay frozen, so the recipe transfers to other latent-diffusion text-to-audio systems.
  • Three-step generation approximately matches a 25-step flow-matching model (LAFMA), indicating that rectified diffusion is competitive with flow matching in the low-step regime.
  • Classifier-free guidance remains beneficial after rectification, with guidance scale 1.5 giving the best objective scores, so the accelerated model can still trade off diversity against text adherence.

Reading between the lines

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

  • Because the training set is entirely teacher-generated, the model cannot exceed the teacher's coverage; a natural stress test is to evaluate on captions describing audio types that are rare in AudioCaps, where a synthetic-only corpus may underperform.
  • The same rectified-diffusion pairing could be applied to other conditional audio tasks (music from text, singing, sound effects) by reusing a pretrained latent diffusion backbone rather than training a flow-matching model from scratch.
  • The paper's stated future direction of one-step generation suggests distillation on top of the rectified trajectories; an intermediate testable step is whether two-step sampling with a consistency model preserves the reported CLAP and OVL margins over LAFMA.
  • Since the training pairs come from a frozen teacher, the recipe is cheap to reproduce; readers should check whether the gains survive when the teacher, dataset, or evaluation package is swapped, since the reported numbers depend on Auffusion's specific output distribution.
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

4 major / 5 minor

Summary. The paper proposes AudioTurbo, a text-to-audio (TTA) generation method that applies rectified diffusion to a pre-trained latent diffusion model (Auffusion). The authors train the UNet on deterministic noise-audio pairs generated by Auffusion from AudioCaps captions, enabling generation in 5-10 sampling steps. Experiments on the AudioCaps test set report that AudioTurbo with 10 steps achieves better FD, KL, IS, CLAP, OVL, and REL than Tango, AudioLDM2, and LAFMA at 200 steps, and that with 3 steps it roughly matches LAFMA's 25-step performance. The central claim is that rectified diffusion, when initialized from a strong pretrained teacher, yields fast and high-quality TTA generation.

Significance. If the reported results were unbiased, the contribution would be practically significant: it demonstrates that rectified diffusion can be adapted to a pretrained TTA model, reducing inference to 10 steps (or even 3) while retaining quality comparable to 200-step diffusion baselines. The method is simple, uses publicly available components, and addresses a real deployment bottleneck of TTA systems. However, the current evaluation has three load-bearing weaknesses: the main comparison omits the teacher model Auffusion, the classifier-free guidance scale is selected on the test set, and all metrics are single-point estimates without error bars. These issues materially affect the validity of the headline claims, so the significance can only be confirmed after the evaluation is strengthened.

major comments (4)
  1. [Table 1 and Section 4.1.3] The main comparison in Table 1 omits Auffusion, even though Section 4.1.3 lists Auffusion as a baseline and the method is built directly from it (the UNet is initialized from Auffusion and all training pairs are generated by Auffusion). Table 3 shows that Auffusion at 200 steps achieves CLAP 30.5, which is higher than AudioTurbo's 29.8 at 10 steps, and Auffusion at 25 steps already reaches 30.2. Since Auffusion is the most relevant comparison, its omission makes the abstract claim that AudioTurbo 'outperforms prior models' misleading. The authors should include Auffusion in Table 1 and qualify the claim per metric (e.g., AudioTurbo improves FD and KL over Auffusion but not CLAP).
  2. [Section 4.2.2 and Table 2] The CFG scale w=1.5 is selected by evaluating metrics on the same AudioCaps test set that is later used for the final results in Table 1 and Table 3. This is a form of test-set leakage that inflates the reported metrics and undermines the unbiasedness of the 'outperforms' claims. The authors should either select the guidance scale on a held-out validation split or report results for all guidance scales on the test set without choosing the best a posteriori.
  3. [Tables 1, 2, and 3] All objective metrics are single-point estimates with no error bars, confidence intervals, or repeated runs. Given the small differences in CLAP (e.g., 29.8 vs 30.5) and the non-monotonic behavior in Table 3 (AudioTurbo's FD at 200 steps is 21.75, slightly worse than 21.69 at 25 steps), it is impossible to assess whether the reported differences are statistically significant. The authors should provide variance estimates over multiple seeds or bootstrap resampling.
  4. [Section 4.2.3 and Conclusion] The claim that AudioTurbo 'achieves comparable performance in just three steps, matching LAFMA's optimal performance' is only partially supported by Table 3. At 3 steps, AudioTurbo has better KL (1.52 vs 1.58) and CLAP (25.4 vs 22.9) than LAFMA at 25 steps, but worse FD (29.96 vs 27.37). Since the metrics disagree, the 'comparable' claim should be made per metric, and the abstract's phrasing 'reduces inference to 3 steps compared to a flow-matching-based acceleration model' should be similarly qualified.
minor comments (5)
  1. [Section 4.1.4] The subjective evaluation uses six evaluators and 50 samples per model; please report the inter-annotator agreement (e.g., ICC or Krippendorff's alpha) and provide confidence intervals, as this sample size is small.
  2. [Table 1] AudioLDM2 is trained on 'AC+4 others' while the other baselines and AudioTurbo are trained on AudioCaps only; this makes the comparison partly uncontrolled. Please state this difference explicitly in the table caption or add a model trained on AudioCaps only.
  3. [Section 2] The derivation in Equations (1)-(6) largely replicates known results from Rectified Diffusion [21] and DPM-Solver [25]. The authors should more clearly state what is a review of prior theory and what is their own contribution, to avoid the impression that a new derivation is being presented.
  4. [Section 4.1.1] The entire training set consists of AudioCaps captions paired with audio generated by Auffusion. The authors should explicitly discuss the potential for teacher bias: the reported test metrics may partly reflect Auffusion's generation distribution, and it would strengthen the paper to test with a small amount of real audio or a different teacher.
  5. [Table 3] AudioTurbo's performance at 200 steps is not better than at 25 steps on any objective metric (e.g., FD 21.75 vs 21.69, KL 1.32 vs 1.31). This non-monotonic trend suggests possible training/sampling instability; a brief comment or an error-bar analysis would help.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity in the derivation; one mild test-set-fitted CFG choice affects the headline numbers.

  1. fitted input called prediction [Section 4.2.2 (Table 2) and Section 4.2.1 (Table 1)]
    "We achieve the best performance across nearly all metrics for AudioTurbo with a guidance scale of 1.5, while the metrics deteriorate as the guidance scale increases further. Adjusting the guidance scale carefully allows us to improve the overall effectiveness of our approach. Therefore, in the other experiments of this paper, we fix the guidance scale of AudioTurbo at 1.5."

    AudioTurbo's headline numbers in Tables 1 and 3 are computed on the AudioCaps test set described in Section 4.1.1 ('For evaluation, we use the test subset of AudioCaps'). Section 4.2.2 selects the CFG scale by comparing FD/KL/IS/CLAP on that same evaluation set, then fixes w=1.5 for all subsequent experiments. Consequently, the reported 'prediction' (e.g., CLAP 29.8 at 10 steps) is not an independent out-of-sample result: the inference hyperparameter was chosen to optimize the very metric on the very set that is later presented as evidence. This is a mild fitted-input-called-prediction loop. It is not a construction-level equivalence: the model could have produced different metrics for the chosen scale, and the derivation in Section 2 is independent of this choice.

full rationale

The paper's central derivation is not circular. Section 2 summarizes rectified diffusion [21] and derives the first-order ODE form; the training procedure in Section 3.2 samples noise, uses Auffusion to generate latent targets, and retrains the UNet on deterministic (epsilon, z0) pairs. This is a teacher-student distillation setup, but the evaluation is against real AudioCaps test audio, so the benchmarks are external and the model can in principle be worse or better than the teacher (indeed Table 3 shows AudioTurbo's CLAP at 10 steps, 29.8, is below Auffusion's 200-step 30.5, while FD/KL improve). No claim in the derivation reduces to its own input by construction. The self-citations to AudioLDM / AudioLDM2 and the audioldm_eval package are baselines and a code-reproduced evaluation harness, not load-bearing justifications. The only circular-adjacent practice is the test-set selection of the CFG scale in Section 4.2.2, which introduces selection bias into the headline numbers and is captured as a single mild step. The omission of Auffusion from Table 1, despite listing it as a baseline in 4.1.3, is a baseline-comparison concern rather than a circularity.

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

The central method rests on the established rectified diffusion theory and on the quality of the pretrained Auffusion teacher. The only explicit free parameter is the CFG scale, which is tuned on the test set. No new physical or conceptual entities are introduced.

free parameters (1)
  • classifier-free guidance scale w = 1.5
    Selected by maximizing objective metrics on the AudioCaps test set (Table 2, Section 4.2.2), then fixed for all other experiments.
assumptions (3)
  • domain assumption Rectified diffusion produces consistent ODE trajectories for deterministic noise-sample pairs
    Assumed from the cited Rectified Diffusion paper [21] in Section 2; the paper does not prove it.
  • domain assumption The pretrained Auffusion model generates audio from captions that is suitable as training data for a student model
    Used in Section 4.1.1 to create 47,744 training samples; if the teacher is weak or narrow, the student inherits those weaknesses.
  • domain assumption AudioCaps captions and the evaluation metrics are adequate proxies for text-to-audio quality
    The paper evaluates on AudioCaps test set with FD, KL, IS, and CLAP; this is standard practice but an assumption about what 'good' means.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AudioTurbo: Fast Text-to-Audio Generation with Rectified Diffusion." pith.science (2026). https://pith.science/paper/PCCIB7KF

@misc{pith2026250522106,
  author       = {Pith},
  title        = {Pith review of: AudioTurbo: Fast Text-to-Audio Generation with Rectified Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PCCIB7KF}},
  note         = {Machine review of arXiv:2505.22106}
}
read the original abstract

Diffusion models have significantly improved the quality and diversity of audio generation but are hindered by slow inference speed. Rectified flow enhances inference speed by learning straight-line ordinary differential equation (ODE) paths. However, this approach requires training a flow-matching model from scratch and tends to perform suboptimally, or even poorly, at low step counts. To address the limitations of rectified flow while leveraging the advantages of advanced pre-trained diffusion models, this study integrates pre-trained models with the rectified diffusion method to improve the efficiency of text-to-audio (TTA) generation. Specifically, we propose AudioTurbo, which learns first-order ODE paths from deterministic noise sample pairs generated by a pre-trained TTA model. Experiments on the AudioCaps dataset demonstrate that our model, with only 10 sampling steps, outperforms prior models and reduces inference to 3 steps compared to a flow-matching-based acceleration model.

Figures

Figures reproduced from arXiv: 2505.22106 by the authors.

Figure 1
Figure 1. An overview of AudioTurbo architecture. Note that the trainable parameters are initialized using the pretrained TTA model, Auffusion. achieves superior performance with significantly fewer sam￾pling steps. Specifically, compared to several advanced base￾line models, AudioTurbo generates higher-quality audio and ex￾hibits better text-audio alignment capabilities in just five steps. 2. Rectified Diffusion Diffusion mo… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FdAudio: MeanFlow-Anchored Fr\'echet-Distance Post-Training for One-Step Text-to-Audio Generation

    eess.AS 2026-07 conditional novelty 5.5 of 10

    MeanFlow-anchored multi-representation FD post-training improves one-step text-to-audio quality without collapsing multi-step sampling.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    The audio generated by TTA systems is primarily categorized into three types: speech, sound effects, and music

    Introduction Text-to-audio (TTA) generation is a task that aims to gen- erate audio samples based on natural language descriptions [1, 2, 3, 4, 5]. The audio generated by TTA systems is primarily categorized into three types: speech, sound effects, and music. These audio outputs have been widely applied in fields such as voice assistants [6, 7], game deve...

    work page

  2. [2]

    Rectified Diffusion Diffusion models operate by gradually adding noise to trans- form a complex data distribution into a known prior distribution and then learning to reverse the process by progressively de- noising the prior to reconstruct the original data distribution. The forward diffusion process can be modeled by a continuous- time stochastic differ...

    work page

  3. [3]

    Proposed Method We propose AudioTurbo, a generative model based on rectified diffusion for efficient TTA generation. AudioTurbo consists of four key components: 1) a text encoder for generating text em- beddings; 2) a LDM based on rectified diffusion for predicting audio latent representations; 3) a V AE decoder that reconstructs the mel-spectrograms; and...

    work page

  4. [4]

    Experimental Setup 4.1.1

    Experiments 4.1. Experimental Setup 4.1.1. Dataset Following previous works [5, 19], we conduct our experiments using the AudioCaps (AC) [31] dataset. It is a publicly avail- able dataset that contains high-quality human-annotated cap- tions. We first use Auffusion along with AudioCaps captions to generate47,744ten-second audio samples for training. For e...

    work page

  5. [5]

    Experiments show that by combining SOTA TTA models with rectified dif- fusion, AudioTurbo surpasses baseline models in both objec- tive and subjective evaluations

    Conclusion We introduced AudioTurbo, a rectified diffusion-based model for TTA, designed to leverage the strengths of SOTA pre-trained TTA models while enhancing inference efficiency. Experiments show that by combining SOTA TTA models with rectified dif- fusion, AudioTurbo surpasses baseline models in both objec- tive and subjective evaluations. Moreover,...

    work page

  6. [6]

    Text-to- speech and speech-to-text converter—voice assistant,

    S. Janokar, S. Ratnaparkhi, M. Rathi, and A. Rathod, “Text-to- speech and speech-to-text converter—voice assistant,” inInven- tive Systems and Control: Proceedings of ICISC 2023. Springer, 2023, pp. 653–664

    work page 2023

  7. [7]

    AudioGen: Textu- ally guided audio generation,

    F. Kreuk, G. Synnaeve, A. Polyak, U. Singer, A. D ´efossez, J. Copet, D. Parikh, Y . Taigman, and Y . Adi, “AudioGen: Textu- ally guided audio generation,” inThe Eleventh International Con- ference on Learning Representations

    work page

  8. [8]

    UniAudio: An audio founda- tion model toward universal audio generation,

    D. Yang, J. Tian, X. Tan, R. Huang, S. Liu, X. Chang, J. Shi, S. Zhao, J. Bian, X. Wuet al., “UniAudio: An audio founda- tion model toward universal audio generation,”arXiv preprint arXiv:2310.00704, 2023

    arXiv 2023

Show all 41 references
  1. [9]

    Diffsound: Discrete diffusion model for text-to-sound genera- tion,

    D. Yang, J. Yu, H. Wang, W. Wang, C. Weng, Y . Zou, and D. Yu, “Diffsound: Discrete diffusion model for text-to-sound genera- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1720–1733, 2023

    2023

  2. [10]

    AudioLDM: text-to-audio generation with latent diffusion models,

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “AudioLDM: text-to-audio generation with latent diffusion models,” inProceedings of the 40th International Conference on Machine Learning, 2023, pp. 21 450–21 474

    2023

  3. [11]

    Text-to- audio generation using instruction guided latent diffusion model,

    D. Ghosal, N. Majumder, A. Mehrish, and S. Poria, “Text-to- audio generation using instruction guided latent diffusion model,” inProceedings of the 31st ACM International Conference on Mul- timedia, 2023, pp. 3590–3598

    2023

  4. [12]

    Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,

    J. Xue, Y . Deng, Y . Gao, and Y . Li, “Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 32, pp. 4700–4712, 2024

    2024

  5. [13]

    A survey of deep learning audio gener- ation methods,

    M. Bo ˇzi´c and M. Horvat, “A survey of deep learning audio gener- ation methods,”arXiv preprint arXiv:2406.00146, 2024

    2024 arXiv

  6. [14]

    Lever- aging AI to generate audio for user-generated content in video games,

    T. Marrinan, P. Akram, O. Gurmessa, and A. Shishkin, “Lever- aging AI to generate audio for user-generated content in video games,”arXiv preprint arXiv:2404.17018, 2024

    2024 arXiv

  7. [15]

    MusicLM: Generating music from text,

    A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi et al., “MusicLM: Generating music from text,”arXiv preprint arXiv:2301.11325, 2023

    2023 arXiv

  8. [16]

    Neural discrete representa- tion learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representa- tion learning,”Advances in Neural Information Processing Sys- tems, vol. 30, 2017

    2017

  9. [17]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

    2022

  10. [18]

    FlashSpeech: Efficient zero-shot speech synthesis,

    Z. Ye, Z. Ju, H. Liu, X. Tan, J. Chen, Y . Lu, P. Sun, J. Pan, W. Bian, S. Heet al., “FlashSpeech: Efficient zero-shot speech synthesis,” inProceedings of the 32nd ACM International Con- ference on Multimedia, 2024, pp. 6998–7007

    2024

  11. [19]

    AudioLDM 2: Learn- ing holistic audio generation with self-supervised pretraining,

    H. Liu, Y . Yuan, X. Liu, X. Mei, Q. Kong, Q. Tian, Y . Wang, W. Wang, Y . Wang, and M. D. Plumbley, “AudioLDM 2: Learn- ing holistic audio generation with self-supervised pretraining,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, 2024

    2024

  12. [20]

    into the audio latent space to enable text-guided audio gen- eration. LAFMA is capable of producing high-quality audio samples by utilizing a numerical ordinary differential equation (ODE) solver, effectively reducing the inference steps to around ten. A drawback of LAFMA is i...

    2025 arXiv

  13. [21]

    Make-an-Audio: Text-to-audio generation with prompt-enhanced diffusion models,

    R. Huang, J. Huang, D. Yang, Y . Ren, L. Liu, M. Li, Z. Ye, J. Liu, X. Yin, and Z. Zhao, “Make-an-Audio: Text-to-audio generation with prompt-enhanced diffusion models,” inInternational Con- ference on Machine Learning. PMLR, 2023, pp. 13 916–13 932

    2023

  14. [22]

    Make-an-Audio 2: Temporal-enhanced text-to-audio generation,

    J. Huang, Y . Ren, R. Huang, D. Yang, Z. Ye, C. Zhang, J. Liu, X. Yin, Z. Ma, and Z. Zhao, “Make-an-Audio 2: Temporal-enhanced text-to-audio generation,”arXiv preprint arXiv:2305.18474, 2023

    2023 arXiv

  15. [23]

    ProDiff: Progressive fast diffusion model for high-quality text-to-speech,

    R. Huang, Z. Zhao, H. Liu, J. Liu, C. Cui, and Y . Ren, “ProDiff: Progressive fast diffusion model for high-quality text-to-speech,” inProceedings of the 30th ACM International Conference on Mul- timedia, 2022, pp. 2595–2605

    2022

  16. [24]

    Matcha-TTS: A fast TTS architecture with conditional flow matching,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-TTS: A fast TTS architecture with conditional flow matching,” inInternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 341–11 345

    2024

  17. [25]

    LAFMA: A latent flow matching model for text-to-audio generation,

    W. Guan, K. Wang, W. Zhou, Y . Wang, F. Deng, H. Wang, L. Li, Q. Hong, and Y . Qin, “LAFMA: A latent flow matching model for text-to-audio generation,”arXiv preprint arXiv:2406.08203, 2024

    2024 arXiv

  18. [26]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh Inter- national Conference on Learning Representations

  19. [27]

    Rectified diffusion: Straightness is not your need in rectified flow,

    F.-Y . Wang, L. Yang, Z. Huang, M. Wang, and H. Li, “Rectified diffusion: Straightness is not your need in rectified flow,”arXiv preprint arXiv:2410.07303, 2024

    2024 arXiv

  20. [28]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  21. [29]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  22. [30]

    Score-based generative modeling through stochas- tic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochas- tic differential equations,” inInternational Conference on Learn- ing Representations, 2021

    2021

  23. [31]

    DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,”Advances in Neural Information Processing Sys- tems, vol. 35, pp. 5775–5787, 2022

    2022

  24. [32]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” inIn- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  25. [33]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning. PMLR, 2021, pp. 8748–8763

    2021

  26. [34]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241

    2015

  27. [35]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representa- tions, 2021

    2021

  28. [36]

    Pseudo numerical methods for diffusion models on manifolds,

    L. Liu, Y . Ren, Z. Lin, and Z. Zhao, “Pseudo numerical methods for diffusion models on manifolds,” inInternational Conference on Learning Representations, 2022

    2022

  29. [37]

    AudioCaps: Generat- ing captions for audios in the wild,

    C. D. Kim, B. Kim, H. Lee, and G. Kim, “AudioCaps: Generat- ing captions for audios in the wild,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Vol- ume 1 (Long and Short Papers...

    2019

  30. [38]

    PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumb- ley, “PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

    2020

  31. [39]

    Diffusion models beat GANs on im- age synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on im- age synthesis,”Advances in Neural Information Processing Sys- tems, vol. 34, pp. 8780–8794, 2021

    2021

  32. [40]

    Universal sound separation with self-supervised audio masked autoencoder,

    J. Zhao, X. Liu, J. Zhao, Y . Yuan, Q. Kong, M. D. Plumbley, and W. Wang, “Universal sound separation with self-supervised audio masked autoencoder,” in32nd European Signal Processing Conference (EUSIPCO). IEEE, 2024, pp. 1–5

    2024

  33. [41]

    SoloAudio: Target sound extraction with language-oriented audio diffusion transformer,

    H. Wang, J. Hai, Y .-J. Lu, K. Thakkar, M. Elhilali, and N. De- hak, “SoloAudio: Target sound extraction with language-oriented audio diffusion transformer,”arXiv preprint arXiv:2409.08425, 2024

    2024 arXiv

Pith tools

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