Pith. sign in

REVIEW 3 major objections 5 minor 51 references

LILAC: An Idempotent Neural Speech Codec

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper proves that a neural speech codec built from an invertible analysis transform and an idempotent scalar quantizer is codec idempotent by construction, and demonstrates it in LILAC at 0.75 kbit/s.

desk verdict A genuinely new construction for idempotent speech codecs, with a clean exact-arithmetic proof and strong empirical verification, but the universal floating-point guarantee has a gap and the affine stem contradicts the no-division claim. read the letter →

arxiv 2608.05727 v1 pith:6DVRKKFL submitted 2026-08-06 cs.SD cs.LG

June Young Yi , Dongwook Lee , Jiheum Yeom , Sungroh Yoon This is my paper · ORCID
classification cs.SDcs.LG
keywords idempotentcodecneuralaudioinvertibletransformfinitescalarquantizationlow-bitratespeechcodingdecode-re-encodedriftstability
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 tries to establish that a neural speech codec can be made codec idempotent—decoding a token stream and re-encoding it returns the exact same stream—not by training losses but by architectural construction. It presents LILAC, a fully convolutional 24 kHz codec at 9.375 Hz and 0.75 kbit/s, whose encoder is an invertible analysis transform that discards selected coordinates, and whose decoder inverts the transform after a learned fill predicts the discarded coordinates. The idempotence guarantee holds for any learned weights and any fill function, provided the finite scalar quantizer is idempotent, which clamping and rounding are. This matters because repeated decode-re-encode cycles in speech pipelines currently degrade most codecs; a codec whose tokens are a fixed point keeps downstream language-model-style systems stable.

What carries the argument

The load-bearing mechanism is the invertible analysis transform: a stack of additive coupling blocks (splitting even/odd time indices, adding a learned function, then subtracting it to invert) interleaved with orthogonal 1×1 convolutions whose inverse is their transpose. These operations are volume-preserving and division-free, so the inverse is exact in floating point. The retained coordinates are quantized by Finite Scalar Quantization (FSQ), a coordinate-wise clamp and round to a fixed grid; being a projection, it is idempotent. The decoder runs the transform in reverse while a convolutional fill network predicts the discarded coordinates, and this fill is arbitrary—the fixed point does not depend on it.

What would settle it

Take the released checkpoint and run a search over valid token streams whose coordinates sit near a FSQ grid boundary: decode each stream, re-encode it, and test whether any quantized coordinate rounds to a different index. A single mismatch—or a discrepancy between full-precision and bf16 decodes of the same token stream—would refute the claimed universal idempotence. The paper’s verification covers 7,457 test clips and 100 cycles, but not all possible streams.

Watch

Extended reading notes

Core claim

LILAC's central claim is a structural fixed-point theorem: for an invertible analysis transform $A(x) = (z, f)$ with retained coordinates $z$ and discarded coordinates $f$, a finite scalar quantizer $q$, and a decoder $D(z_q) = A^{-1}(z_q, \varphi(z_q))$ using any learned fill $\varphi$, the encoder $E(x) = q(\Pi_z A(x))$ satisfies $E(D(z_q)) = z_q$ for every $z_q$ in the image of $q$, regardless of weights. The proof runs: $A(D(z_q)) = (z_q, \varphi(z_q))$, so selecting the retained coordinates gives back $z_q$; because clamping and rounding are idempotent, quantizing $z_q$ again returns $z_q$. The paper verifies the property on all 7,457 LibriSpeech and LibriTTS-R test samples and over 100 cycles, and reports UTMOS of 4.14 and 4.24 on those sets at 0.75 kbit/s.

Load-bearing premise

The decoder must implement the exact inverse of the analysis transform in floating point, not just in real arithmetic; the paper explicitly pins only the 1×1 convolutions to full precision, while bf16 matmul and a learned-scale division in the affine stem could, on some token stream, round a retained coordinate across a quantization boundary and break $E(D(z_q)) = z_q$.

Editorial extensions

If this is right

  • Any number of decode-re-encode cycles leaves LILAC's token stream bit-identical, so every quality metric (UTMOS, dWER, speaker EER) is constant across cycles; the paper shows this on 100-cycle runs.
  • The idempotence proof is weight-independent and applies to a stochastic decoder $\varphi(z_q, \omega)$, so LILAC can be extended to diverse outputs without losing the fixed-point property.
  • Because the fixed point is structural rather than trained, the same construction transfers to other bitrates, frame rates, or audio types by adjusting the number of retained coordinates and the FSQ grid.
  • Existing codecs rewrite at least 15% of tokens on a single pass; LILAC is presented as the first neural speech codec where decode-re-encode drift is eliminated by design rather than reduced by losses.

Reading between the lines

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

  • The proof never uses training data, so the same invertible-transform construction should yield idempotent codecs for music or general audio with no architectural change beyond retraining the fill.
  • A practical consequence the paper leaves implicit: with a guaranteed fixed point, a downstream language model could cache or edit tokens without fear that re-encoding an unchanged region will silently rewrite neighboring tokens—an assumption most speech-LLM pipelines cannot currently make.
  • The weakest spot is numerical: since the guarantee is exact-arithmetic, a stress test that pushes FSQ boundaries under bf16 matmul, or on hardware with different rounding, would decide whether the real-world implementation inherits the paper's universal claim.
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 / 5 minor

Summary. LILAC is a fully convolutional 24 kHz speech codec operating at 0.75 kbit/s and 9.375 Hz that claims codec idempotence by construction: re-encoding the decoded audio of any valid token stream returns the identical token stream. The encoder is an invertible analysis transform (orthogonal 1x1 convolutions plus additive/affine coupling) followed by finite scalar quantization that retains a subset of coordinates; the decoder predicts discarded coordinates and inverts the transform. The paper gives an exact-arithmetic proof of idempotence, verifies it empirically on 7,457 test clips and 100 decode-re-encode cycles, and reports quality comparisons against sub-1 kbit/s baselines, including a crowdsourced listening test.

Significance. The idempotence-by-construction idea is novel and potentially valuable for speech-generation pipelines that store, edit, or retransmit codec tokens. The central proposition is a clean, weight-independent derivation from the architecture, and the empirical verification (7,457 samples, 100 cycles, all bit-identical) is unusually thorough. The paper also reports seed variance, multiple datasets, a human listening test with confidence intervals, and public code/checkpoints, which are strengths. However, the manuscript's universal claim for the released implementation is not fully substantiated because the exact-arithmetic proof does not cover floating-point rounding in the implemented inverse transform. If the claim is properly scoped or the implementation is made bit-exact, the contribution would be solid; as written, the central claim is somewhat overstated.

major comments (3)
  1. [Codec Idempotence (Proposition) and Training Details] The Proposition proves idempotence in exact real arithmetic, but the implementation is floating-point. Training Details states that only the invertible 1x1 convolution layers are executed in the full-precision setting, while other layers use bf16 matmul precision. The decoder applies the inverse of the analysis transform in floating point; additive coupling inversion (e.g., x1 = y2 - g(y1)) is not bit-exact under round-to-nearest, because rounding in g(y1) or in the subtraction can change the low-order bits. If such a perturbation moves a retained coordinate across a quantization boundary, E(D(z_q)) = z_q can fail for some token stream that is not in the 7,457 test clips. The empirical coverage does not establish the universal 'any valid token stream' claim. Please either (i) prove that floating-point rounding in every layer of the implemented inverse cannot shift a retained coordinate across a quantization boundary (e.g., a margin argument that holds for all weights and all grid points), or (ii) implement the inverse so that it is bit-exact (e.g., integer arithmetic or full precision throughout), or (iii) explicitly restrict the formal claim to exact arithmetic and describe the empirical result as verification on audio-derived streams only.
  2. [Analysis Transform (Stem stage)] The affine stem uses y2 = s·x2 + b with learned scale s and bias b. Its inverse requires division: x2 = (y2 - b)/s. This directly contradicts the Background statement that 'the inverse transform requires no division. This prevents floating-point overflows.' Division by a learned, potentially small s is not exactly invertible with multiplication in floating point and can overflow or underflow. The same affine operation also contradicts the 'volume-preserving transform' claim, since scaling by s changes the volume by the product of the s entries. Please either replace the affine stem with a division-free invertible operation (e.g., additive coupling without scaling) or revise the no-division and volume-preserving statements to acknowledge the exception.
  3. [Abstract and Results (Idempotence Verification)] The abstract claims that 're-encoding the decoded audio of any valid token stream returns the identical stream' for LILAC as a system. The verification in Results covers all 7,457 LibriSpeech/LibriTTS-R test samples and 100 cycles on 100 clips. The token space is astronomically large, and the tested streams are natural-speech encodings, a tiny subset. If 'valid token stream' means any element of the image of the quantizer, then a simple random-sampling test over the 80-bit token grid would provide much stronger evidence (or find a counterexample). Please either add such a test and report the outcome, or narrow the claim to 'all tested audio-derived token streams' unless the floating-point issue in the first major comment is resolved.
minor comments (5)
  1. [Figure 2 caption] The phrase 'closing the exact red loop for any fill' is unclear; please specify what the red loop represents in the figure.
  2. [Supplement Table 6 and main text] The dataset name is written as 'HiFiTTS-2' everywhere except Table 6, where it appears as 'HiFi-TTS 2'; please make the spelling consistent.
  3. [Supplement Table 24] The 'Duplicate-reference disagreement' values (median 4.5, mean 9.25) lack units; please state that these are MUSHRA points.
  4. [Training Details] The 'full-precision setting' for invertible 1x1 convolutions is not specified (fp32 vs fp64); please clarify the exact precision used.
  5. [Limitations] The Limitations section acknowledges that downstream benefits are theoretical, but it does not mention the floating-point exactness caveat for the idempotence guarantee; adding this caveat would make the limitation statement more complete.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: codec idempotence is derived algebraically from the invertible-transform construction and verified independently.

full rationale

LILAC's central claim is the Proposition that for every z_q in the image of q and every fill φ, E(D(z_q)) = z_q holds regardless of weights. The proof is a direct expansion of the definitions: E(x) = q(Π_z A(x)) and D(z_q) = A^{-1}(z_q, φ(z_q)). Since A is stated to be invertible, A(D(z_q)) = (z_q, φ(z_q)); projection gives z_q; and finite scalar quantization is clamping and rounding, both idempotent. No parameter is fitted to make this equation true, no subset of test data is used to tune the guarantee, and no self-citation supplies the load-bearing invertibility premise. The learned fill φ is explicitly allowed to be arbitrary and hence cannot smuggle in the result. The empirical verification on 7,457 LibriSpeech/LibriTTS-R samples and 100-cycle checks is a sanity check of the implementation, not the source of the claim. The manuscript's own limitation statements, such as 'the anticipated benefits of codec idempotence ... are only theoretical,' are candid and weaken the downstream utility claim without creating circular support. The floating-point caveat raised by the reviewer—bf16 matmul precision outside the full-precision 1×1 convolutions and division in the affine stem—concerns whether the implemented decoder exactly realizes the exact-arithmetic inverse, which is a correctness/robustness issue rather than a circularity in the derivation. Accordingly, no circular step is present.

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

The idempotence guarantee depends on no fitted constants: the proof is algebraic and holds for any weights, any fill, and any bit allocation. Quality-related hyperparameters, such as loss weights, bit allocation, and frame rate, are design choices rather than parameters fitted to make the guarantee true. The ledger instead records the structural assumptions about exact invertibility, shared weights, quantizer idempotence, and dataset separation that the central claim and its evaluation rest on.

assumptions (4)
  • domain assumption The analysis transform A is exactly invertible, including orthogonal 1x1 convolutions, additive couplings, and the affine stem coupling with nonzero scale.
    The idempotence proof uses A^{-1}; the paper builds A from invertible primitives but does not formally prove bit-exact inversion for every floating-point operator.
  • standard math Finite scalar quantization q = clamp + round is idempotent on its image.
    Used in the proof E(D(z_q)) = q(z_q) = z_q; true for exact arithmetic and for re-quantizing already-quantized coordinates.
  • domain assumption The encoder and decoder share the same analysis transform weights, so the decoder's inverse matches the encoder's forward.
    Required for E(D(z_q)) to be the projection of A on the waveform generated by A^{-1}; any weight mismatch would break the argument.
  • domain assumption Training on HiFiTTS-2 with reader-filtering makes LibriSpeech, LibriTTS-R, and VCTK valid out-of-domain evaluations.
    Quality comparisons depend on this; VCTK is corpus-level disjoint, while LibriSpeech and LibriTTS-R share the LibriVox domain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LILAC: An Idempotent Neural Speech Codec." pith.science (2026). https://pith.science/paper/6DVRKKFL

@misc{pith2026260805727,
  author       = {Pith},
  title        = {Pith review of: LILAC: An Idempotent Neural Speech Codec},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DVRKKFL}},
  note         = {Machine review of arXiv:2608.05727}
}
read the original abstract

Neural Audio Codecs are widely adopted in speech generation and editing. However, existing neural audio codecs are not idempotent: across the paper's twelve baseline systems, every configuration tested rewrites, on average, at least 15% of its tokens in a single decode-re-encode pass. This poses a problem for utilizing Neural Audio Codecs as token interfaces in pipelines where re-encoding decoded outputs can occur. We present LILAC, a fully convolutional 24 kHz speech codec at 9.375 Hz and 0.75 kbit/s that is codec idempotent by construction; re-encoding the decoded audio of any valid token stream returns the identical stream. LILAC achieves idempotency while maintaining competitive quality, reaching UTMOS 4.14 and 4.24 on LibriSpeech and LibriTTS-R test sets, comparable to SOTA sub-1 kbit/s Neural Audio Codecs.

Figures

Figures reproduced from arXiv: 2608.05727 by the authors.

Figure 1
Figure 1. Iterated re-encoding on 100 reader-balanced LibriSpeech test-clean clips: token agreement with the first encoding, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. LILAC’s encoder (top), shared-weight inverse (middle), and decoder-only fill networks (bottom). Projection stages [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. dWER versus released bitrate for the sub-1 kb/s panel; lines connect operating points from the same checkpoint family. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Speaker EER and dWER after each codec self-encodes its own output. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    Zeghidour, Neil and Luebs, Alejandro and Omran, Ahmed and Skoglund, Jan and Tagliasacchi, Marco , journal =

    work page

  2. [2]

    Transactions on Machine Learning Research , year =

    High Fidelity Neural Audio Compression , author =. Transactions on Machine Learning Research , year =

    work page

  3. [3]

    High-Fidelity Audio Compression with Improved

    Kumar, Rithesh and Seetharaman, Prem and Luebs, Alejandro and Kumar, Ishaan and Kumar, Kundan , booktitle =. High-Fidelity Audio Compression with Improved

    work page

  4. [4]

    Zhang, Xin and Zhang, Dong and Li, Shimin and Zhou, Yaqian and Qiu, Xipeng , booktitle =

  5. [5]

    2024 , eprint =

    Moshi: A Speech-Text Foundation Model for Real-Time Dialogue , author =. 2024 , eprint =

    work page 2024

  6. [6]

    Della Libera, Luca and Paissan, Francesco and Subakan, Cem and Ravanelli, Mirco , booktitle =

    work page

  7. [7]

    2409.05377 , archivePrefix =

    Xin, Detai and Tan, Xu and Takamichi, Shinnosuke and Saruwatari, Hiroshi , year =. 2409.05377 , archivePrefix =

    arXiv

  8. [8]

    2024 , eprint =

    Siuzdak, Hubert and Gr. 2024 , eprint =

    work page 2024

Show all 51 references
  1. [9]

    Ji, Shengpeng and Jiang, Ziyue and Wang, Wen and Chen, Yifu and Fang, Minghui and Zuo, Jialong and Yang, Qian and Cheng, Xize and Wang, Zehan and Li, Ruiqi and others , booktitle =

  2. [10]

    Low Frame-Rate Speech Codec: A Codec Designed for Fast High-Quality Speech

    Casanova, Edresson and Langman, Ryan and Neekhara, Paarth and Hussain, Shehzeen and Li, Jason and Ghosh, Subhankar and Juki. Low Frame-Rate Speech Codec: A Codec Designed for Fast High-Quality Speech. IEEE International Conference on Acoustics, Speech and Signal Processing (IC...

  3. [11]

    1410.8516 , archivePrefix =

    Dinh, Laurent and Krueger, David and Bengio, Yoshua , year =. 1410.8516 , archivePrefix =

    arXiv

  4. [12]

    Density Estimation Using

    Dinh, Laurent and Sohl-Dickstein, Jascha and Bengio, Samy , booktitle =. Density Estimation Using

  5. [13]

    and Dhariwal, Prafulla , booktitle =

    Kingma, Diederik P. and Dhariwal, Prafulla , booktitle =

  6. [14]

    Journal of Machine Learning Research , volume =

    Normalizing Flows for Probabilistic Modeling and Inference , author =. Journal of Machine Learning Research , volume =

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Neural Discrete Representation Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  8. [16]

    Finite Scalar Quantization:

    Mentzer, Fabian and Minnen, David and Agustsson, Eirikur and Tschannen, Michael , booktitle =. Finite Scalar Quantization:

  9. [17]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    The Perception-Distortion Tradeoff , author =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  10. [18]

    International Conference on Machine Learning (ICML) , year =

    Rethinking Lossy Compression: The Rate-Distortion-Perception Tradeoff , author =. International Conference on Machine Learning (ICML) , year =

  11. [19]

    International Conference on Learning Representations (ICLR) , year =

    Idempotent Generative Network , author =. International Conference on Learning Representations (ICLR) , year =

  12. [20]

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

    Code Drift: Towards Idempotent Neural Audio Codecs , author =. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

  13. [21]

    Multirate Systems and Filter Banks , author =

  14. [22]

    Liu, Zhuang and Mao, Hanzi and Wu, Chao-Yuan and Feichtenhofer, Christoph and Darrell, Trevor and Xie, Saining , booktitle =. A

  15. [23]

    Kong, Jungil and Kim, Jaehyeon and Bae, Jaekyoung , booktitle =

  16. [24]

    arXiv preprint arXiv:1308.3432 , year =

    Estimating or Propagating Gradients through Stochastic Neurons for Conditional Computation , author =. arXiv preprint arXiv:1308.3432 , year =

    arXiv

  17. [25]

    International Conference on Machine Learning (ICML) , year =

    Robust Speech Recognition via Large-Scale Weak Supervision , author =. International Conference on Machine Learning (ICML) , year =

  18. [26]

    Saeki, Takaaki and Xin, Detai and Nakata, Wataru and Koriyama, Tomoki and Takamichi, Shinnosuke and Saruwatari, Hiroshi , booktitle =

  19. [27]

    and Beerends, John G

    Rix, Antony W. and Beerends, John G. and Hollier, Michael P. and Hekstra, Andries P. , booktitle =. Perceptual Evaluation of Speech Quality (

  20. [28]

    IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , year =

    A Short-Time Objective Intelligibility Measure for Time-Frequency Weighted Noisy Speech , author =. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , year =

  21. [29]

    , booktitle =

    Le Roux, Jonathan and Wisdom, Scott and Erdogan, Hakan and Hershey, John R. , booktitle =

  22. [30]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Neural Networks Fail to Learn Periodic Functions and How to Fix It , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  23. [31]

    Fundamentals of Statistical Signal Processing: Estimation Theory , author =

  24. [32]

    Elements of Information Theory , author =

  25. [33]

    Librispeech: An

    Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev , booktitle =. Librispeech: An

  26. [34]

    Koizumi, Yuma and Zen, Heiga and Karita, Shigeki and Ding, Yifan and Yatabe, Kohei and Morioka, Nobuyuki and Bacchiani, Michiel and Zhang, Yu and Han, Wei and Bapna, Ankur , booktitle =

  27. [35]

    Yamagishi, Junichi and Veaux, Christophe and MacDonald, Kirsten , howpublished =

  28. [36]

    Richter, Julius and Wu, Yi-Chiao and Krenn, Steven and Welker, Simon and Lay, Bunlong and Watanabe, Shinji and Richard, Alexander and Gerkmann, Timo , booktitle =

  29. [37]

    Langman, Ryan and Yang, Xuesong and Neekhara, Paarth and Hussain, Shehzeen and Casanova, Edresson and Bakhturina, Evelina and Li, Jason , booktitle =

  30. [38]

    Du, Zhihao and Zhang, Shiliang and Hu, Kai and Zheng, Siqi , booktitle =

  31. [39]

    Yang, Dongchao and Liu, Songxiang and Huang, Rongjie and Tian, Jinchuan and Weng, Chao and Zou, Yuexian , journal =

  32. [40]

    Li, Jiaqi and Lin, Xiaolong and Li, Zhekai and Huang, Shixi and Wang, Yuancheng and Wang, Chaoren and Zhan, Zhenpeng and Wu, Zhizheng , booktitle =

  33. [41]

    Ragano, Alessandro and Skoglund, Jan and Hines, Andrew , booktitle =

  34. [42]

    Conference on Uncertainty in Artificial Intelligence (UAI) , year =

    Averaging Weights Leads to Wider Optima and Better Generalization , author =. Conference on Uncertainty in Artificial Intelligence (UAI) , year =

  35. [43]

    SIAM Journal on Mathematical Analysis , volume =

    Sweldens, Wim , title =. SIAM Journal on Mathematical Analysis , volume =

  36. [44]

    IEEE Spoken Language Technology Workshop (SLT) , pages =

    Yoshimura, Takenori and Hashimoto, Kei and Oura, Keiichiro and Nankaku, Yoshihiko and Tokuda, Keiichi , title =. IEEE Spoken Language Technology Workshop (SLT) , pages =

  37. [45]

    Advances in Neural Information Processing Systems , year =

    Li, Yanghao and Xu, Tongda and Wang, Yan and Liu, Jingjing and Zhang, Ya-Qin , title =. Advances in Neural Information Processing Systems , year =

  38. [46]

    Advances in Neural Information Processing Systems , year =

    The Reversible Residual Network: Backpropagation Without Storing Activations , author =. Advances in Neural Information Processing Systems , year =

  39. [47]

    International Conference on Learning Representations , year =

    Reformer: The Efficient Transformer , author =. International Conference on Learning Representations , year =

  40. [48]

    Chen, Sanyuan and Wang, Chengyi and Chen, Zhengyang and Wu, Yu and Liu, Shujie and Chen, Zhuo and Li, Jinyu and Kanda, Naoyuki and Yoshioka, Takuya and Xiao, Xiong and Wu, Jian and Zhou, Long and Ren, Shuo and Qian, Yanmin and Qian, Yao and Wu, Jian and Zeng, Michael and Yu, X...

  41. [49]

    IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , year =

    Miipher: A Robust Speech Restoration Model Integrating Self-Supervised Speech and Text Representations , author =. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , year =

  42. [50]

    International Conference on Learning Representations , year =

    Image and Video Tokenization with Binary Spherical Quantization , author =. International Conference on Learning Representations , year =

  43. [51]

    International Conference on Learning Representations , year =

    Decoupled Weight Decay Regularization , author =. International Conference on Learning Representations , year =

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.