Pith. sign in

REVIEW 3 major objections 6 minor 15 references

ZeroSep: Separate Anything in Audio with Zero Training

T0 review · 3 major / 6 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 A pre-trained text-to-audio diffusion model, used with a null inversion prompt and a guidance weight of one, separates mixed audio into described sources with zero training.

desk verdict A clever zero-training separation recipe built on diffusion inversion, but the evaluation does not yet show it recovers the actual source rather than a plausible class-consistent sound. read the letter →

arxiv 2505.23625 v1 pith:Y3WQIVZW submitted 2025-05-29 cs.SD cs.CVeess.AS

Chao Huang , Yuesheng Ma , Junxuan Huang , Susan Liang , Yunlong Tang , Jing Bi , Wenqiang Liu , Nima Mesgarani
show 1 more author
Chenliang Xu
This is my paper
classification cs.SDcs.CVeess.AS
keywords zero-shotaudiosourceseparationtext-guideddiffusioninversionDDIMDDPMclassifier-freeguidancetraining-freeopen-set
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

The paper claims that audio source separation — isolating one sound from a mixture — can be done with zero task-specific training by a pre-trained text-to-audio diffusion model used out of the box. The recipe is to invert the mixed audio into the model's latent noise space, then denoise from that point under a text prompt naming the source to extract, with the classifier-free guidance weight set to $\omega = 1$. With that configuration the authors report separation quality that matches or beats supervised systems on the MUSIC and AVE benchmarks, including the best training-based language-queried separator. If correct, high-quality open-set separation no longer needs labeled mixtures, and every improvement in generative audio models becomes an improvement in separation.

What carries the argument

The load-bearing object is the pair formed by a latent inversion operator and the classifier-free guidance rule. The inversion operator $F_{\text{inv}}(x_0, c)$, realized as DDIM inversion (Eq. (2)) or DDPM inversion with per-step noise vectors (Eqs. (3)–(5)), maps the mixture's VAE-encoded mel-spectrogram $x_{\text{mix}}$ to a noise latent $x_T$. The reverse process then applies $\epsilon_t = \epsilon_\theta(x_t, \varnothing, t) + \omega\,(\epsilon_\theta(x_t, c_{\text{rev}}, t) - \epsilon_\theta(x_t, \varnothing, t))$, where the choice $\omega = 1$ cancels the unconditional term and leaves the text-conditioned noise prediction as the sole steering signal. That cancellation is the mechanism that flips the model from generating new audio to isolating the source described by $c_{\text{rev}}$; the ablations show $\omega > 1$ degrades separation below the mixture baseline, and a reverse prompt that does not name the actual target collapses separation quality.

What would settle it

Take the two-source mixtures from MUSIC, run ZeroSep with the target prompt, and measure the residual energy of the other source in each output as a signal-to-interference ratio (SIR) against the ground-truth interferer. If SIR is near zero or negative across many samples, the method is producing prompt-consistent audio rather than isolating sources; the paper's own Table 3 (a VAE-vocoder reconstruction already loses about 23 dB of SDR) and Appendix B (interferers that survive separation) show this outcome is live.

Watch

Extended reading notes

Core claim

ZeroSep's discovery is that a pre-trained text-guided audio diffusion model becomes a zero-shot source separator through configuration alone, with no fine-tuning and no separation-specific data. The authors identify three conditions that convert synthesis into separation: the inversion prompt is a null prompt $c_{\text{inv}} = \varnothing$ (or a description of the non-target sources), the reverse denoising prompt is the text description of the target source $c_{\text{rev}} = c^{(i)}$, and the classifier-free guidance weight is $\omega = 1$ rather than the large values used for generation. At $\omega = 1$ the combined noise prediction of Eq. (6) collapses to the conditional prediction $\epsilon_\theta(x_t, c_{\text{rev}}, t)$, so each denoising step follows the target source's textual direction instead of the unconditional mixture reconstruction that $\omega = 0$ would produce. The same pipeline works across several pre-trained backbones (AudioLDM, AudioLDM2, Tango) and with both DDIM and DDPM inversion, and the paper reports that stronger generators yield better separators. The headline quantitative result is that on MUSIC, ZeroSep outperforms the leading supervised baseline FlowSep on all four reported metrics (FAD, LPAPS, CLAP-A, CLAP-T), while on AVE it is comparable.

Load-bearing premise

Inverting the mixture into latent noise must preserve each source's actual content well enough that text-conditioned denoising recovers the true recorded source, rather than generating new audio that merely fits the prompt.

Editorial extensions

If this is right

  • Separation performance tracks generation performance: across AudioLDM-S/L, AudioLDM2-S/L, and Tango, better generative FAD correlates with better separation metrics, so a stronger text-to-audio generator directly yields a stronger separator.
  • Open-set separation becomes native to the method, since any source a user can describe in text is a separation target with no closed vocabulary or retraining.
  • The pipeline is backbone-agnostic and inversion-agnostic: AudioLDM, AudioLDM2, and Tango all separate successfully, and DDPM inversion gives more stable scores than DDIM across model sizes.
  • On MUSIC, the training-free ZeroSep beats the strongest supervised language-queried baseline FlowSep on all four reported metrics, and on AVE it matches it, so skipping task-specific training costs nothing on these benchmarks.
  • Because the guidance weight can be swept continuously, the same mechanism interpolates from full mixture reconstruction at $\omega = 0$ to complete separation at $\omega = 1$, which the authors connect to sound-highlighting applications.

Reading between the lines

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

  • If the $\omega = 1$ cancellation is the true mechanism, the same invert-then-denoise-with-a-prompt recipe should transfer to other conditional diffusion domains (images, video) as zero-shot 'separate what the prompt names' — a testable extension the paper does not run.
  • The paper's move to embedding-space metrics is forced rather than cosmetic: Table 3 shows the VAE-vocoder reconstruction path alone costs about 23 dB of SDR, so what is promised is perceptual and semantic fidelity, not waveform fidelity; downstream usability of the outputs is the open question.
  • A natural repair for the documented failure mode of Appendix B — high-energy interferers surviving one pass — is to feed the output back through inversion and re-denoise or increase inference steps, since the paper postulates interference removal is incremental.
  • The generation-quality/separation-quality correlation implies that separation state of the art advances as a by-product of generative audio state of the art, potentially shifting investment away from bespoke separation corpora.
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 / 6 minor

Summary. The paper proposes ZeroSep, a training-free audio source separation method that repurposes pre-trained text-guided latent audio diffusion models. The pipeline consists of inverting the latent encoding of an audio mixture via DDIM or DDPM inversion to obtain a noise trajectory, then denoising from that latent with classifier-free guidance weight omega=1 under a text prompt describing the target source. The authors evaluate on MUSIC and AVE using FAD, LPAPS, CLAP-A, and CLAP-T metrics, compare with supervised baselines (LASS-Net, AudioSep, FlowSep) and training-free baselines (NMF-MFCC, AudioEdit), and report that ZeroSep matches or exceeds FlowSep on several metrics. Ablations study the influence of the backbone model, inversion method, guidance weight, and prompt choices, and an appendix documents a failure case. The paper claims that the method achieves zero-shot separation without any task-specific training and that stronger generative models improve separation quality.

Significance. If the central claim were established, namely that the denoised output is the actual source instance from the mixture rather than a new audio from the same class, ZeroSep would be a notable conceptual contribution: it would show that a pre-trained text-to-audio diffusion model can be inverted into a discriminative separator, with open-set capability through text prompts and no task-specific data. The paper's strengths include experiments across multiple backbones (AudioLDM, AudioLDM2, Tango), two inversion techniques, an explicit failure-case appendix, and a public demo page. However, the current evaluation uses only distributional and semantic embedding metrics, and the paper itself concedes inversion approximation errors in Section 3.1 and lists latent inversion artifacts as a limitation. Because no instance-level fidelity measurement is provided, the significance is conditional on additional evidence that separation, and not class-conditional generation, is occurring.

major comments (3)
  1. [Sec. 4.1 and Table 3] The decision to omit SDR/SIR/SAR is not adequately justified. Table 3 shows that a VAE-vocoder reconstruction of the mixture has SDR -23.07 dB relative to the target source while the original mixture itself has SDR 0.31 dB; this demonstrates that generative outputs can have low SDR even when they carry mixture content, but it does not show that ZeroSep's outputs carry the target source content. The central claim that separated sources are highly faithful to the original sources (Section 1) requires a reference-based check that the output is the specific source instance. I recommend reporting SDR/SIR/SAR or SI-SDR on at least a subset, together with alignment or a known-instance identification test (for example, speaker or instrument identity), to rule out the alternative explanation that omega=1 simply generates a new audio consistent with the prompt.
  2. [Sec. 4.1 and Table 1] The chosen metrics cannot distinguish source-instance recovery from class-conditional generation. CLAP-T is the cosine similarity between the output audio and the text embedding of the target class, which is closely related to the objective that a text-to-audio model is trained to maximize; FAD is a set-level distribution distance; and CLAP-A operates in a learned space that is largely invariant to within-class instance identity. High scores on these metrics are compatible with generating any plausible example of the prompt class from the inverted latent. A human listening study asking whether the particular source (for example, the same cello phrase or the same speaker) is recovered, or a retrieval-based identity test, would substantially strengthen the separation claim.
  3. [Sec. 3.3, Eq. (6), and Appendix B] The proposed mechanism is consistent with the alternative hypothesis that the method performs prompt-conditioned regeneration rather than separation. With omega=1 in Eq. (6), the noise prediction is purely conditional on crev, so the reverse process is identical to text-to-audio sampling except for the starting latent obtained by inversion. The authors acknowledge in Section 3.1 that inversion accumulates errors, and Appendix B documents a mixture (man speech plus shofar) where the interfering source is not removed; these observations are exactly what one would expect if the latent carried only coarse content and the output was generated to match the prompt. To support the separation interpretation, the paper should include a diagnostic that varies the identity of the mixture while holding the prompt fixed and shows that the output tracks the mixture's instance-level content, for example by comparing outputs from two different cello performances with the same prompt.
minor comments (6)
  1. [Sec. 2 and Sec. 4.1] The baseline is called AudioEdit in the related work and tables but AudioEditor in Section 4.1; please use a consistent name.
  2. [Table 4] The caption uses the same symbol for improvement and degradation; presumably one of the triangles should be an upward-pointing triangle.
  3. [Sec. 4.2 and Table 1] The claim that ZeroSep surpasses supervised methods is too strong for AVE, where ZeroSep's CLAP-A (0.442) and CLAP-T (-0.001) are below FlowSep's (0.493 and 0.082); consider qualifying this as competitive rather than superior.
  4. [Appendix B] The failure-case discussion is qualitative; adding a quantitative measure, such as the residual energy of the interfering source, would clarify the frequency and severity of such failures.
  5. [Eqs. (1)-(2)] The notation for the noise schedule alpha-bar is not defined before first use; please add a definition or a reference.
  6. [Fig. 3(b)-(c)] The claimed positive correlation between generative FAD and separation metrics is presented without statistical significance or confidence intervals; please quantify the trend if it is used as evidence for scaling behavior.

Circularity Check

1 steps flagged · score 3.0 of 10

ZeroSep is an empirical recipe with no derivation-level circularity, but its central evidence partly reduces to the base model's generation objective: with ω=1 the reverse pass is ordinary text-conditioned generation, and CLAP-T scores measure the same text-audio alignment the pretrained model was trained to maximize.

  1. self definitional [Section 3.3, Eq. (6) and Section 4.1, Evaluation Metrics]
    "Setting ω = 1 removes the unconditional noise estimation from the combined prediction in Eq. (6), leaving only the component aligned with the target source described by crev. ... CLAP-T is the cosine similarity between audio embeddings and the text embedding of the target class."

    With ω=1, Eq. (6) reduces to ϵt = ϵθ(xt, crev, t), which is exactly the ordinary classifier-free text-conditioned generation step. The pretrained audio diffusion model (AudioLDM2, Tango, etc.) was trained so that this step produces audio matching crev, and CLAP-T is precisely the cosine similarity between the output audio embedding and the crev text embedding. High CLAP-T therefore verifies prompt-following by construction and does not establish that the output is the particular source instance present in the mixture. The paper's own Tab. 3 shows that the VAE-vocoder path loses waveform identity (SDR −23 dB), so the semantic/distributional metrics remain confounded with class-conditional generation.

full rationale

ZeroSep is an empirical method, not a mathematical derivation, so the severe circularity modes do not directly apply. The inversion operator Finv and the denoising sampler are standard operations from external work (DDIM/DDPM inversion; AudioLDM/AudioLDM2/Tango), and the text prompts are dataset labels rather than quantities fitted to the test outputs. The guidance weight ω is selected on the same benchmarks, which is a correctness or overfitting concern, but not a circular reduction. The only genuinely self-referential piece of evidence is CLAP-T: with ω=1, Eq. (6) collapses to ϵt = ϵθ(xt, crev, t), i.e., ordinary conditional generation, and CLAP-T measures exactly the text-audio alignment that this pretrained generative model was optimized to produce. Thus the high CLAP-T scores in Tables 1 and 2 are, by construction, a measure of prompt-following rather than evidence that the specific source instance was recovered from the mixture. The paper's decision to omit SDR/SIR/SAR (Tab. 3) reinforces this: the VAE-vocoder path loses waveform identity, leaving only distributional and semantic metrics that a class-conditional generator can satisfy. This is an evaluation-level circularity, not a derivation-level one, so the score is 3 rather than higher. Self-citations in the paper (e.g., DAVIS splits, future-application discussion) are not load-bearing for the central claim.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its load-bearing assumptions are: the inversion math is correct, the pretrained text-to-audio model can be repurposed for extraction, the VAE-vocoder pipeline preserves content, and the chosen perceptual metrics measure separation. The main free choices are the guidance weight, inversion prompt, and backbone, all selected empirically on the evaluation benchmarks.

free parameters (3)
  • classifier-free guidance weight omega = 1
    Empirically selected from the sweep {0, 0.5, 1, 1.5, 2} in Figure 3(a); the paper states it yields the best separation metrics.
  • inversion prompt cinv = null prompt (empty)
    Chosen as the default in Section 3.3; Table 4 shows alternatives such as c(j) give similar or slightly different results, so the choice is empirical.
  • base diffusion backbone and inversion method = AudioLDM2-Large with DDIM inversion
    The main reported ZeroSep row in Table 1 is the best configuration from Table 2; the backbone and inversion method were selected using the same benchmark datasets.
assumptions (4)
  • standard math DDIM and DDPM inversion equations correctly map a clean latent to a noise trajectory that can reconstruct the input (Sections 3.1, equations 1 through 5).
    The paper relies on Song et al. and Huberman-Spiegelglas et al. for the inversion formulas and does not rederive them.
  • domain assumption A pretrained text-guided audio diffusion model can generate audio that aligns with a text prompt, and this alignment transfers to a discriminative separation setting (Section 3.3).
    The entire method assumes the conditional denoiser, when guided at omega=1, selects the source described by the prompt rather than producing unrelated content.
  • domain assumption The VAE latent and vocoder preserve enough source information for separation, even though Table 3 shows very poor waveform-level reconstruction (SDR -23 dB) for the mixture.
    ZeroSep operates in the latent space and outputs waveforms through VAE decoding and a vocoder; if content is lost in this pipeline, the separated output cannot be faithful.
  • domain assumption FAD, LPAPS, and CLAP similarities are valid proxies for separation quality (Section 4.1).
    The paper uses these perceptual and semantic metrics as the main evidence because traditional SDR metrics are deemed misleading for generative outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZeroSep: Separate Anything in Audio with Zero Training." pith.science (2026). https://pith.science/paper/Y3WQIVZW

@misc{pith2026250523625,
  author       = {Pith},
  title        = {Pith review of: ZeroSep: Separate Anything in Audio with Zero Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y3WQIVZW}},
  note         = {Machine review of arXiv:2505.23625}
}
read the original abstract

Audio source separation is fundamental for machines to understand complex acoustic environments and underpins numerous audio applications. Current supervised deep learning approaches, while powerful, are limited by the need for extensive, task-specific labeled data and struggle to generalize to the immense variability and open-set nature of real-world acoustic scenes. Inspired by the success of generative foundation models, we investigate whether pre-trained text-guided audio diffusion models can overcome these limitations. We make a surprising discovery: zero-shot source separation can be achieved purely through a pre-trained text-guided audio diffusion model under the right configuration. Our method, named ZeroSep, works by inverting the mixed audio into the diffusion model's latent space and then using text conditioning to guide the denoising process to recover individual sources. Without any task-specific training or fine-tuning, ZeroSep repurposes the generative diffusion model for a discriminative separation task and inherently supports open-set scenarios through its rich textual priors. ZeroSep is compatible with a variety of pre-trained text-guided audio diffusion backbones and delivers strong separation performance on multiple separation benchmarks, surpassing even supervised methods.

Figures

Figures reproduced from arXiv: 2505.23625 by the authors.

Figure 1
Figure 1. The overview of ZeroSep, which includes (a) an inversion process to obtain a latent representation for the mixture, and (b) a separation denoising process to effectively extract the target source with text conditions. We show the choice of inversion prompt cinv and reverse prompt crev in (c), and demonstrate the valid separation region defined by ω in (d). DDPM Inversion. In contrast to DDIM inversion, DDPM inversio… view at source ↗
Figure 2
Figure 2. Qualitative visualization of audio separation results. The figure shows the input mixture [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (a) Impact of guidance weight ω: increasing ω from 0 to 1 improves separation metrics (LPAPS and CLAP-A), whereas ω > 1 degrades performance below the mixture baseline (ω = 0), underscoring the critical role of ω. (b)–(c) Positive correlation between separation quality (normalized all scores from Tab. 2) and generative capability (normalized FAD scores on AudioCap [Liu et al., 2023b], [Liu et al., 2024]) across Audi… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Failure case analysis of ZeroSep. Mixture: Man speech (stem 1) + Shofar (stem 2). B Failure Case Analysis While generally effective, ZeroSep can sometimes fail to fully isolate the target source. This typically occurs when an interfering source possesses significant en…
Figure 5
Figure 5. Figure 5: Mixture: Cello (stem 1) + Erhu (Stem 2) 0 20 40 60 GT stem 1 0 20 40 60 LASS stem 1 0 20 40 60 FLOW stem 1 0 20 40 60 EDIT stem 1 0 20 40 60 SEP stem 1 0 20 40 60 OURS stem 1 0 20 40 60 GT stem 2 0 20 40 60 LASS stem 2 0 20 40 60 FLOW stem 2 0 20 40 60 EDIT stem 2 0 20…
Figure 7
Figure 7. Figure 7: Mixture: Accordion (stem 1) + Flute (Stem 2) 0 20 40 60 GT stem 1 0 20 40 60 LASS stem 1 0 20 40 60 FLOW stem 1 0 20 40 60 EDIT stem 1 0 20 40 60 SEP stem 1 0 20 40 60 OURS stem 1 0 20 40 60 GT stem 2 0 20 40 60 LASS stem 2 0 20 40 60 FLOW stem 2 0 20 40 60 EDIT stem 2…
Figure 9
Figure 9. Figure 9: Mixture: Tuba (stem 1) + Cello (Stem 2) 0 20 40 60 GT stem 1 0 20 40 60 LASS stem 1 0 20 40 60 FLOW stem 1 0 20 40 60 EDIT stem 1 0 20 40 60 SEP stem 1 0 20 40 60 OURS stem 1 0 20 40 60 GT stem 2 0 20 40 60 LASS stem 2 0 20 40 60 FLOW stem 2 0 20 40 60 EDIT stem 2 0 20…
Figure 11
Figure 11. Figure 11: Mixture: Truck (stem 1) + Banjo (Stem 2) 0 20 40 60 GT stem 1 0 20 40 60 LASS stem 1 0 20 40 60 FLOW stem 1 0 20 40 60 EDIT stem 1 0 20 40 60 SEP stem 1 0 20 40 60 OURS stem 1 0 20 40 60 GT stem 2 0 20 40 60 LASS stem 2 0 20 40 60 FLOW stem 2 0 20 40 60 EDIT stem 2 0 …
Figure 14
Figure 14. Figure 14: Mixture: Male Speech (stem 1) + Airplane (Stem 2) 18 [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Mixture: Truck (stem 1) + Ukulele (Stem 2) 0 20 40 60 GT stem 1 0 20 40 60 LASS stem 1 0 20 40 60 FLOW stem 1 0 20 40 60 EDIT stem 1 0 20 40 60 SEP stem 1 0 20 40 60 OURS stem 1 0 20 40 60 GT stem 2 0 20 40 60 LASS stem 2 0 20 40 60 FLOW stem 2 0 20 40 60 EDIT stem 2 …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 6 canonical work pages

  1. [4]

    Weiss, Mohammad Norouzi, and William Chan

    Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, and William Chan. Wave- grad: Estimating gradients for waveform generation. In International Conference on Learning Representations (ICLR), 2021a. Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-tts: A denoising diffusion model for text-to-speech. In Pr...

    work page 2021

  2. [5]

    Vadim Popov, Ivan V ovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov

    doi: 10.21437/Interspeech.2021-469. Vadim Popov, Ivan V ovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599–8608. PMLR,

    work page doi:10.21437/interspeech.2021-469 2021

  3. [7]

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D

    doi: 10.21437/Interspeech.2021-36. Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D. Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In Proceedings of the 40th International Conference on Machine Learning (ICML), pages 21450–21474. PMLR, 2023b. Haohe Liu, Yi Yuan, Xubo Liu, Xinhao Mei, ...

    work page doi:10.21437/interspeech.2021-36 2021

  4. [8]

    Deepanway Ghosal, Navonil Majumder, Ambuj Mehrish, and Soujanya Poria

    doi: 10.1109/TASLP.2024.3399607. Deepanway Ghosal, Navonil Majumder, Ambuj Mehrish, and Soujanya Poria. Text-to-audio gener- ation using instruction tuned llm and latent diffusion model. arXiv preprint arXiv:2304.13731,

    arXiv 2024

  5. [10]

    URL https://proceedings.mlr.press/v235/manor24a. html. Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502,

    arXiv 2010

  6. [12]

    Listen, think, and understand

    Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlinsky, and James Glass. Listen, think, and understand. arXiv preprint arXiv:2305.10790,

    arXiv

  7. [14]

    Av-nerf: Learning neural fields for real-world audio-visual scene synthesis

    Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36:37472–37490, 2023a. Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse re...

    arXiv 2025

  8. [2014]

    Fr\’echet audio distance: A metric for evaluating music enhancement algorithms

    Kevin Kilgour, Mauricio Zuluaga, Dominik Roblek, and Matthew Sharifi. Fr\’echet audio distance: A metric for evaluating music enhancement algorithms. arXiv preprint arXiv:1812.08466,

    arXiv

Show all 15 references
  1. [2017]

    Chao Huang, Susan Liang, Yapeng Tian, Anurag Kumar, and Chenliang Xu

    doi: 10.1109/ICASSP.2017.7952154. Chao Huang, Susan Liang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. High-quality visually- guided sound separation from diverse categories. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 35–49, December 2024a. Yuanch...

    2017

  2. [2019]

    Cem Subakan, Mirco Ravanelli, Samuele Cornell, Mirko Bronzi, and Jianyuan Zhong

    doi: 10.1109/TASLP.2019.2915167. Cem Subakan, Mirco Ravanelli, Samuele Cornell, Mirko Bronzi, and Jianyuan Zhong. Attention is all you need in speech separation. In Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),

    2019

  3. [2021]

    Plumbley, and Wenwu Wang

    Xubo Liu, Haohe Liu, Qiuqiang Kong, Xinhao Mei, Jinzheng Zhao, Qiushi Huang, Mark D. Plumbley, and Wenwu Wang. Separate what you describe: Language-queried audio source separation. In Proc. Interspeech, pages 1801–1805, 2022a. Xubo Liu, Qiuqiang Kong, Yan Zhao, Haohe Liu, Yi Y...

    arXiv

  4. [2022]

    Diffsinger: Singing voice synthesis via shallow diffusion mechanism

    Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022b. Junhyeok Lee and Seungu Han. Nu-wave: A diffusion probabilistic model for...

    2021

  5. [2023]

    Gama: A large audio-language model with advanced audio understanding and complex reasoning abilities

    Sreyan Ghosh, Sonal Kumar, Ashish Seth, Chandra Kiran Reddy Evuru, Utkarsh Tyagi, S Sakshi, Oriol Nieto, Ramani Duraiswami, and Dinesh Manocha. Gama: A large audio-language model with advanced audio understanding and complex reasoning abilities. arXiv preprint arXiv:2406.11768,

    arXiv

  6. [2024]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,

    1901

  7. [2025]

    Concept sliders: Lora adaptors for precise control in diffusion models

    Rohit Gandikota, Joanna Materzy´nska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. InEuropean Conference on Computer Vision, pages 172–188. Springer, 2024a. Chao Huang, Susan Liang, Yunlong Tang, Yapeng ...

    arXiv

Pith tools

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