Pith. sign in

REVIEW 3 major objections 4 minor 40 references

Post-training a one-step text-to-audio model with multi-encoder Fréchet distance, anchored by MeanFlow consistency, raises one-step quality to match multi-step systems without destroying multi-step sampling.

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

2026-07-14 11:51 UTC pith:C2TVYCHG

load-bearing objection Clean diagnosis of FD post-training multi-step collapse, fixed with a MeanFlow anchor; real one-step gains on MeanAudio, but absolute SOTA numbers sit on a thin one-epoch 80k run. the 3 major comments →

arxiv 2607.10421 v1 pith:C2TVYCHG submitted 2026-07-11 eess.AS cs.SD

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

Kuan-Po Huang , Bo-Ru Lu , Ho-Lam Chung , Shih-Hsin Wang , Hung-yi Lee This is my paper
classification eess.AS cs.SD
keywords text-to-audio generationone-step generationMeanFlowFréchet distanceflow matchingfew-step samplingpost-training
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.

Few-step text-to-audio models can generate sound in a single network pass by learning average velocities, yet that one-step output still trails the quality of slow multi-step diffusion or flow models. This paper shows that a short post-training stage that directly matches the one-step output distribution to real audio across several frozen embedding spaces (via Fréchet distance) closes most of that gap. Alone, the same Fréchet objective collapses the model into a pure one-step mapper and ruins multi-step sampling; adding a lightweight MeanFlow consistency term as an anchor keeps the progressive velocity field intact. The resulting 120 M model therefore delivers state-of-the-art one-step scores among few-step systems while a 25-step path remains competitive with far larger multi-step generators at a fraction of their latency. A sympathetic reader cares because the method gives both real-time single-pass generation and an optional high-fidelity multi-step mode from one set of weights.

Core claim

Naive Fréchet-distance post-training improves one-step text-to-audio quality but overwrites the learned velocity field, so multi-step sampling degrades sharply. Anchoring that distribution-matching loss with a MeanFlow consistency objective on intermediate intervals restores trajectory consistency, yielding a single model whose one-step outputs reach state-of-the-art few-step quality and whose multi-step path stays high-fidelity.

What carries the argument

The MeanFlow-anchored multi-representation Fréchet loss: L_total = L_Multi-FD + λ L_MF, where L_Multi-FD matches one-step generated statistics to real-data statistics in an ensemble of frozen audio encoders (via a queue estimator) and L_MF enforces average-velocity consistency on random sub-intervals so the same weights remain integrable for multi-step sampling.

Load-bearing premise

That one epoch of post-training on a random 80 000-clip subset, with a fixed four-encoder ensemble and a single λ chosen by ablation, produces a generally robust recipe rather than a fit tuned to the AudioCaps test distribution.

What would settle it

Retrain the identical pipeline on a held-out audio-caption corpus never seen during the 80 k post-training stage and measure whether the one-step FD/FAD gains and the 25-step non-collapse both disappear relative to the unanchored MeanAudio baseline.

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

If this is right

  • One-step text-to-audio generation can reach quality levels previously reserved for multi-step systems while still finishing in under two seconds for a batch of eight 10-second clips.
  • A single 120 M checkpoint can serve both real-time one-step applications and higher-fidelity 25-step applications without separate models or retraining.
  • Distribution-level post-training across multiple frozen audio encoders becomes a practical alternative to distillation or adversarial fine-tuning for few-step generators.
  • The same MeanFlow anchor can be expected to protect multi-step capability whenever any terminal-only distribution loss is applied to a flow-based audio generator.

Where Pith is reading between the lines

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

  • The same dual-stream recipe (terminal FD plus trajectory regularizer) should transfer to other continuous generative modalities that already possess a MeanFlow-style average-velocity objective.
  • Excluding the evaluation encoder (VGGish) from the training ensemble is a general safeguard against metric gaming that future Fréchet-style post-training papers will likely adopt.
  • If the queue-based FD estimator remains stable at larger scales, the method could be applied as a cheap final polishing stage after any large-scale flow-matching pre-training run.

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

3 major / 4 minor

Summary. The paper proposes FdAudio, a post-training procedure for MeanFlow-based text-to-audio generators. Starting from a pretrained MeanAudio checkpoint, it jointly optimizes a multi-representation Fréchet-distance loss (over PANNs, PaSST, BEATs and AudioMAE embeddings, with a queue-based estimator) on one-step samples together with a MeanFlow consistency regularizer on intermediate intervals. The FD term is intended to improve the one-step output distribution while the MeanFlow term is claimed to preserve the velocity field so that the same weights remain usable for multi-step sampling. On AudioCaps the method reports state-of-the-art one-step metrics among few-step systems (FD 12.71, FAD 1.26) and competitive 25-step numbers, with ablations on the anchor weight λ and encoder combinations.

Significance. If the empirical findings hold under stronger experimental controls, the work supplies a practical and conceptually clean recipe for applying distribution-matching post-training to flow-based audio models without destroying multi-step integrability—an issue the authors correctly diagnose and ablate. The dual-stream objective, the multi-encoder FD formulation, the public code/demo, and the clear λ=0 collapse demonstration are concrete contributions to the few-step T2A literature. The limited data regime and single-run reporting, however, currently leave the absolute SOTA claim only partially substantiated.

major comments (3)
  1. [IV.A / Table I] Section IV.A and Table I: the SOTA one-step claim (11.4 % FD / 28.8 % FAD relative to MeanAudio-S-Full) rests on a single post-training run of one epoch on a random 80 k subset of AudioCaps+WavCaps. No multi-seed statistics, no full-data counterpart, and no error bars are reported. Because the absolute numbers are the central quantitative claim, this experimental thinness is load-bearing and must be strengthened (at least multi-seed means ± std or a full-data ablation).
  2. [IV.C / Eq. (8)] Section IV.C and the multi-representation FD loss (Eq. 8): EMA checkpoints are selected by best validation FD, and the training ensemble includes PANNs—the same backbone used for the primary reported FD metric—while only VGGish is excluded. This creates a partial circularity risk for the headline FD numbers even though FAD itself is not directly optimized. The paper should either remove PANNs from the training ensemble for the main results or report an independent held-out embedding metric with confidence intervals.
  3. [V.B / Table III] Table III (λ ablation) convincingly shows multi-step collapse when the MeanFlow anchor is removed, yet the paper offers no further diagnostic that the velocity field itself remains a valid average-velocity field after joint training (e.g., trajectory-wise reconstruction error or intermediate-step FD). Without such evidence the claim that “a single set of weights” preserves high-fidelity multi-step sampling rests solely on end-point metrics.
minor comments (4)
  1. [Table II] Subjective evaluation (Table II) uses only 9 raters and 90 clips; confidence intervals are wide and the absolute gains are modest. A larger listening test or clearer statistical testing would strengthen the perceptual claims.
  2. [Fig. 1] Figure 1 caption and surrounding text could more explicitly label the two computational streams (one-step FD vs. intermediate MeanFlow) so that the dual-stream design is immediately clear to a reader skimming the method.
  3. [References / IV.C] Several 2025–2026 arXiv references appear; ensure final camera-ready versions or DOIs are substituted where available, and that the MeanAudio baseline checkpoint hash is pinned for exact reproducibility.
  4. [Eq. (3)] Notation for the adaptive ℓ2 loss ∥·∥a and the stop-gradient operator is introduced late; a short reminder in Eq. (3) would help readers unfamiliar with the MeanFlow paper.

Circularity Check

1 steps flagged

Mild metric-tuning circularity from optimizing multi-rep FD (incl. PANNs) and selecting EMA by validation FD, while reporting FD_PANN gains; FAD/CLAP/subjective remain independent.

specific steps
  1. fitted input called prediction [IV.C Training; Eq. (8); Table I one-step FD]
    "For the multi-representation FD-loss in Eq. (8) we use four complementary pretrained audio encoders: PANNs [27], PaSST [28], BEATs [29], and AudioMAE [30]... We maintain an exponential moving average (EMA) of the model weights and use the EMA weights for all evaluations, selecting the checkpoint with the best validation FD score."

    L_Multi-FD includes the PANNs Fréchet term that defines the reported FD metric; the EMA checkpoint is then chosen by best validation FD. The 11.4% FD reduction (14.35→12.71) versus MeanAudio is therefore statistically encouraged by the training objective and selection procedure rather than an independent out-of-sample prediction. (FAD is protected by excluding VGGish, so the circularity is partial.)

full rationale

This is an empirical post-training methods paper, not a first-principles derivation. The core construction (joint L_Multi-FD + λ L_MF) is self-contained: FD post-training is adapted from an external citation [18] (no author overlap), MeanFlow is reused from the MeanAudio baseline [17]/[31] as an explicit regularizer rather than a uniqueness claim, and the multi-step collapse without the anchor is demonstrated by direct ablation (Table III, λ=0). No self-definitional equations, no uniqueness theorems imported from the authors, no ansatz smuggled via self-citation, and no renaming of known results. The only mild circularity is that the primary reported metric (FD via PANNs) is both a term inside the multi-representation training loss and the criterion for EMA checkpoint selection; the claimed 11.4% FD reduction is therefore partly encouraged by construction. This is standard ML practice and is mitigated by (a) explicit exclusion of VGGish so FAD is not optimized, (b) additional independent metrics (KL, IS, CLAP, subjective REL/OVL), and (c) the multi-step preservation result, which is not forced by the FD term. Score 2 reflects one non-load-bearing fitted-metric step; the central claims retain independent empirical content.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard flow/MeanFlow identities, the Gaussian Fréchet approximation in frozen audio embeddings, a queue-based FD estimator, and several hand-chosen post-training hyperparameters and data subsampling choices. No new physical entities are introduced; the invented pieces are algorithmic constructs (multi-rep FD loss, MeanFlow anchor usage).

free parameters (4)
  • MeanFlow anchor weight λ = 0.25
    Balancing coefficient in L_total = L_Multi-FD + λ L_MF; ablated over {0, 0.25, 0.5, 1.0} and set to 0.25 as optimal.
  • FD queue size N = 20000
    Size of the representation queue used to estimate generated mean/covariance for FD; fixed at 20,000.
  • Multi-FD stability constant c and encoder weights w_j = c=0.01, w_j=1
    Dynamic normalization of per-encoder FD terms; c=0.01 and uniform w_j=1 chosen by authors.
  • Post-training learning rate / batch / epoch / data subsample = 1e-5, batch 2, 80k/1 epoch
    AdamW LR 1e-5, batch 2, one epoch on random 80k of 387k AudioCaps+WavCaps clips; budget-driven choices that affect reported gains.
axioms (4)
  • domain assumption Embeddings of real and generated audio are adequately modeled as multivariate Gaussians so Wasserstein-2 reduces to the Fréchet formula used as training loss.
    Section III-A defines FD via empirical mean/covariance and uses it as L_FD / L_Multi-FD.
  • standard math MeanFlow identity (average velocity equals instantaneous velocity minus (t-r) times total derivative) remains a valid structural constraint during FD post-training.
    Eqs. (2)–(3) and Section III-B; used as the trajectory anchor.
  • domain assumption Frozen pretrained audio encoders (PANNs, PaSST, BEATs, AudioMAE) provide complementary perceptual spaces whose joint FD matching improves generation quality without overfitting a single metric.
    Section III-A multi-representation FD and Table IV ablations.
  • ad hoc to paper One-step FD optimization plus intermediate MeanFlow regression on independently sampled (r,t) is sufficient to preserve multi-step integrability of the same weights.
    Core design claim of Section III-B and Fig. 1 dual-stream training.
invented entities (2)
  • FdAudio joint objective (multi-rep FD + MeanFlow anchor) no independent evidence
    purpose: Simultaneously match one-step output distributions in embedding space and preserve continuous velocity-field structure for multi-step sampling.
    Defined in Eq. (9) and Section III; algorithmic construct rather than a physical entity.
  • Queue-based multi-representation FD estimator for audio post-training no independent evidence
    purpose: Make distribution-level FD gradients tractable without full-batch covariance or unrolled multi-step backprop.
    Adapted from image FD post-training [18] and specialized with audio encoders in III-A.

pith-pipeline@v1.1.0-grok45 · 17852 in / 3297 out tokens · 31746 ms · 2026-07-14T11:51:08.889303+00:00 · methodology

0 comments
read the original abstract

While recent few-step sampling text-to-audio generation models like MeanAudio substantially accelerate generation by modeling average velocities, their strict one-step generation quality still lags significantly behind multi-step counterparts. We propose FdAudio to bridge this gap. Unlike MeanAudio, which relies solely on regression against target velocity fields, our post-training approach optimizes the final one-step distribution directly across pre-trained embedding spaces via a multi-representation Fr\'echet-distance (FD) loss. Crucially, to prevent the multi-step degradation that naive post-training with FD-loss causes, we introduce a MeanFlow consistency objective as a structural anchor. Results demonstrate that FdAudio establishes state-of-the-art one-step T2A generation quality among few-step systems, yielding an 11.4% reduction in FD score and a 28.8% improvement in FAD score relative to the baseline MeanAudio framework. Notably, we solve FD post-training's naive multi-step degradation issue by proposing the MeanFlow anchor, enabling a 25-step sampling path to maintain high-fidelity audio synthesis that matches or surpasses strong multi-step models at a fraction of their computational latency.

Figures

Figures reproduced from arXiv: 2607.10421 by Bo-Ru Lu, Ho-Lam Chung, Hung-yi Lee, Kuan-Po Huang, Shih-Hsin Wang.

Figure 1
Figure 1. Figure 1: Overview of the post-training framework for FdAudio. (a) FD-Loss calculation pipeline: Ground truth audio statistics ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] 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

40 extracted references · 12 linked inside Pith

  1. [1]

    Diffused responsibility: Analyzing the energy consumption of gen- erative text-to-audio diffusion models,

    R. Passoni, F. Ronchini, L. Comanducci, R. Serizel, and F. Antonacci, “Diffused responsibility: Analyzing the energy consumption of gen- erative text-to-audio diffusion models,” in2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2025, pp. 1–5

    2025

  2. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  3. [3]

    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 International Con- ference on Learning Representations, 2023

    2023

  4. [4]

    Audioldm 2: Learning 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: Learning holistic audio generation with self-supervised pretraining,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2871–2883, 2024

    2024

  5. [5]

    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

    Pith/arXiv arXiv 2023

  6. [6]

    Tango 2: Aligning diffusion-based text-to-audio generations through direct preference optimization,

    N. Majumder, C.-Y . Hung, D. Ghosal, W.-N. Hsu, R. Mihalcea, and S. Poria, “Tango 2: Aligning diffusion-based text-to-audio generations through direct preference optimization,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 564–572

    2024

  7. [7]

    Tangoflux: Super fast and faithful text to audio generation with flow matching and clap-ranked preference optimization,

    C.-Y . Hung, N. Majumder, Z. Kong, A. Mehrish, A. A. Bagherzadeh, C. Li, R. Valle, B. Catanzaro, and S. Poria, “Tangoflux: Super fast and faithful text to audio generation with flow matching and clap-ranked preference optimization,”arXiv preprint arXiv:2412.21037, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Impact: Iterative mask- based parallel decoding for text-to-audio generation with diffusion mod- eling,

    K.-P. Huang, S.-w. Yang, H. PHAN, B.-R. Lu, B. Kim, S. Macha, Q. Tang, S. Ghosh, H.-y. Lee, C.-C. Kaoet al., “Impact: Iterative mask- based parallel decoding for text-to-audio generation with diffusion mod- eling,” inF orty-second International Conference on Machine Learning, 2025

    2025

  9. [9]

    Generative audio language modeling with continuous-valued tokens and masked next- token prediction,

    S.-w. Yang, B. Kim, K.-P. Huang, Q. Tang, H. PHAN, B.-R. Lu, H. Sundar, S. Ghosh, H.-y. Lee, C.-C. Kaoet al., “Generative audio language modeling with continuous-valued tokens and masked next- token prediction,” inF orty-second International Conference on Machine Learning, 2025

    2025

  10. [10]

    EzAudio: Enhancing Text-to-Audio Generation with Efficient Diffusion Transformer,

    J. Hai, Y . Xu, H. Zhang, C. Li, H. Wang, M. Elhilali, and D. Yu, “EzAudio: Enhancing Text-to-Audio Generation with Efficient Diffusion Transformer,” inInterspeech 2025, 2025, pp. 4233–4237

    2025

  11. [11]

    Resonate: Reinforcing text-to-audio generation via online feedback from large audio language models,

    X. Li, J. Liu, W. Chen, H. Zhu, Z. Ma, and X. Chen, “Resonate: Reinforcing text-to-audio generation via online feedback from large audio language models,”arXiv preprint arXiv:2603.11661, 2026

    arXiv 2026

  12. [12]

    Consisten- cyTTA: Accelerating Diffusion-Based Text-to-Audio Generation with Consistency Distillation,

    Y . Bai, T. Dang, D. Tran, K. Koishida, and S. Sojoudi, “Consisten- cyTTA: Accelerating Diffusion-Based Text-to-Audio Generation with Consistency Distillation,” inInterspeech 2024, 2024, pp. 3285–3289

    2024

  13. [13]

    Audiolcm: Text-to-audio generation with latent consistency models,

    H. Liu, R. Huang, Y . Liu, H. Cao, J. Wang, X. Cheng, S. Zheng, and Z. Zhao, “Audiolcm: Text-to-audio generation with latent consistency models,”arXiv preprint arXiv:2406.00356, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Audioturbo: Fast text-to-audio generation with rectified diffusion,

    J. Zhao, J. Zhao, H. Liu, Y . Chen, L. Han, X. Liu, M. Plumbley, and W. Wang, “Audioturbo: Fast text-to-audio generation with rectified diffusion,”arXiv preprint arXiv:2505.22106, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Soundctm: Uniting score-based and consistency models for text-to-sound generation,

    K. Saito, D. Kim, T. Shibuya, C.-H. Lai, Z. Zhong, Y . Takida, and Y . Mitsufuji, “Soundctm: Uniting score-based and consistency models for text-to-sound generation,” inAudio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024

    2024

  16. [16]

    Fast text-to-audio generation with one-step sampling via energy-scoring and auxiliary contextual representation distillation,

    K.-P. Huang, B.-R. Lu, B. Kim, M. Lee, Z. Fabian, R. Korzeniowski, Q. Tang, G. V . Steeg, H.-y. Lee, C.-C. Kaoet al., “Fast text-to-audio generation with one-step sampling via energy-scoring and auxiliary contextual representation distillation,”arXiv preprint arXiv:2605.00329, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Meanaudio: Fast and faithful text-to-audio generation with mean flows,

    X. Li, J. Liu, Y . Liang, Z. Niu, W. Chen, and X. Chen, “Meanaudio: Fast and faithful text-to-audio generation with mean flows,”arXiv preprint arXiv:2508.06098, 2025

    arXiv 2025

  18. [18]

    Representation fr ´echet loss for visual generation,

    J. Yang, Z. Geng, X. Ju, Y . Tian, and Y . Wang, “Representation fr ´echet loss for visual generation,”arXiv preprint arXiv:2604.28190, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    One-step latent-free image generation with pixel mean flows,

    Y . Lu, S. Lu, Q. Sun, H. Zhao, Z. Jiang, X. Wang, T. Li, Z. Geng, and K. He, “One-step latent-free image generation with pixel mean flows,” arXiv preprint arXiv:2601.22158, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Improved mean flows: On the challenges of fastforward generative models,

    Z. Geng, Y . Lu, Z. Wu, E. Shechtman, J. Z. Kolter, and K. He, “Improved mean flows: On the challenges of fastforward generative models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 30 467–30 476

    2026

  21. [21]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2818–2826

    2016

  22. [22]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 976–11 986

    2022

  23. [23]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 000–16 009

    2022

  24. [24]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features,

    M. Tschannen, A. Gritsenko, X. Wang, M. F. Naeem, I. Alabdul- mohsin, N. Parthasarathy, T. Evans, L. Beyer, Y . Xia, B. Mustafa et al., “Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features,”arXiv preprint arXiv:2502.14786, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  27. [27]

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

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “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

  28. [28]

    Effi- cient training of audio transformers with patchout,

    K. Koutini, J. Schl ¨uter, H. Eghbal-Zadeh, and G. Widmer, “Effi- cient training of audio transformers with patchout,”arXiv preprint arXiv:2110.05069, 2021

    Pith/arXiv arXiv 2021

  29. [29]

    Beats: Audio pre-training with acoustic tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, and F. Wei, “Beats: Audio pre-training with acoustic tokenizers,”arXiv preprint arXiv:2212.09058, 2022

    Pith/arXiv arXiv 2022

  30. [30]

    Masked autoencoders that listen,

    P.-Y . Huang, H. Xu, J. Li, A. Baevski, M. Auli, W. Galuba, F. Metze, and C. Feichtenhofer, “Masked autoencoders that listen,”Advances in neural information processing systems, vol. 35, pp. 28 708–28 720, 2022

    2022

  31. [31]

    Mean flows for one- step generative modeling,

    Z. Geng, M. Deng, X. Bai, Z. Kolter, and K. He, “Mean flows for one- step generative modeling,”Advances in Neural Information Processing Systems, vol. 38, pp. 75 460–75 482, 2026

    2026

  32. [32]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 32 211–32 252

    2023

  33. [33]

    Classifier-free diffusion guidance,

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

    Pith/arXiv arXiv 2022

  34. [34]

    Cnn architectures for large-scale audio classification,

    S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seyboldet al., “Cnn architectures for large-scale audio classification,” in2017 ieee international conference on acoustics, speech and signal processing (icassp). IEEE, 2017, pp. 131–135

    2017

  35. [35]

    Fr ´echet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr ´echet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms,” inInterspeech 2019, 2019, pp. 2350–2354

    2019

  36. [36]

    Audiocaps: Generating captions for audios in the wild,

    C. D. Kim, B. Kim, H. Lee, and G. Kim, “Audiocaps: Generating 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, V olume 1 (Long and Short Papers), 2019, pp. 119–132

    2019

  37. [37]

    Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research,

    X. Mei, C. Meng, H. Liu, Q. Kong, T. Ko, C. Zhao, M. D. Plumbley, Y . Zou, and W. Wang, “Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 3339–3354, 2024

    2024

  38. [38]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,”Advances in neural information processing systems, vol. 29, 2016

    2016

  39. [39]

    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,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  40. [40]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahmaet al., “Scaling instruction-finetuned language models,”Journal of Machine Learning Research, vol. 25, no. 70, pp. 1–53, 2024

    2024