Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Discrete Speech Unit Extraction via Independent Component Analysis

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Whitening and independent component analysis, applied before k-means clustering, improve the accuracy and compression of discrete speech units extracted from self-supervised speech models.

desk verdict Useful new preprocessing comparison for DSU extraction, but the headline CER gains rest on single runs and need noise bars before they should be believed. read the letter →

arxiv 2501.06562 v1 pith:VWU2QO5B submitted 2025-01-11 eess.AS cs.AIcs.LGcs.SD

Tomohiko Nakamura , Kwanghee Choi , Keigo Hojo , Yoshiaki Bando , Satoru Fukayama , Shinji Watanabe This is my paper · ORCID
classification eess.AScs.AIcs.LGcs.SD
keywords discretespeechunitsindependentcomponentanalysisk-meansclusteringself-supervisedmodelsautomaticrecognitionwhiteningtokenizationlinearpreprocessing
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

Discrete speech units, the token-like clusters obtained by k-means on self-supervised speech representations, are usually extracted straight from raw model features with no attempt to transform the features before clustering. This paper asks whether simple, unsupervised linear preprocessing can make those units better, and shows for the XLS-R-300M model that whitening and independent component analysis (ICA) before k-means improve downstream character error rates while lowering the bit-rate of the unit stream. The best combination, ICA followed by cosine-distance k-means, cuts test-clean character error rate from 3.2% to 2.5% and bit-rate from 394 to 372 bit/s relative to plain Euclidean k-means. The paper also reports that ICA components are more orthogonal than raw cluster centroids and that individual components align with phonetic contrasts such as voicing and place of articulation. These are cheap, unsupervised changes to a widely used recipe, so the result matters for any system that consumes discrete speech tokens.

What carries the argument

The load-bearing mechanism is the observation that k-means clustering is not invariant under linear transformations of its input, so the space in which clustering happens can be changed deliberately. The central object is a preprocessing chain: standardization subtracts per-dimension means and scales to unit variance; PCA rotates the mean-centered data so dimensions are uncorrelated; whitening additionally equalizes variances; and ICA estimates a demixing matrix $W$ by maximum likelihood with a Laplace prior, outputting $X^{(\mathrm{ica})} = X^{(\mathrm{whiten})} W^\top$ so that the resulting components are approximately statistically independent. Applied before k-means with cosine distance, whitening and ICA normalize feature norms and spread the cluster centroids more orthogonally, which is what the paper connects to the improved discrete units.

What would settle it

Run the same DSU extraction and ASR pipeline with, say, ten random seeds for both plain Euclidean k-means and ICA-plus-cosine k-means on XLS-R-300M, using the paper's 5% k-means sample and 100-epoch training; if the test-clean character error rate distributions overlap substantially, the claimed improvement is not established.

Watch

Extended reading notes

Core claim

On the layer-17 representations of XLS-R-300M, the paper argues that standardization, whitening, and ICA each serve as useful preprocessing for k-means clustering, while PCA alone does not because Euclidean k-means is invariant to the orthogonal rotations PCA applies and because transformer representations are anisotropic. Whitening and ICA outperform the others, especially when k-means uses cosine distance rather than Euclidean, because they normalize feature norms. With ICA plus cosine k-means, the paper reports character error rates of 2.5% on the English read-speech test-clean set and 20.9% on the multilingual 1-hour test set, versus 3.2% and 22.9% for plain Euclidean k-means, at a lower bit-rate. The paper is careful to note that this benefit is not universal: on Wav2vec2.0-Large the same preprocessing helps on English test sets but hurts on the multilingual set, and on mHuBERT and WavLM it performs worse than Euclidean k-means, indicating that effectiveness depends on the training methodology of the speech model. Qualitatively, ICA components reveal interpretable phonetic axes, such as voicing contrasts like [B]/[P] and allophonic pairs like [HH]/[HV].

Load-bearing premise

The claimed gains rest on single runs of a stochastic end-to-end ASR training with no repeated seeds, so the reported error-rate differences could be run-to-run noise.

Editorial extensions

If this is right

  • For XLS-R-300M, replacing raw Euclidean k-means with ICA-plus-cosine preprocessing yields better ASR character error rates and a lower bit-rate in the discrete-unit stream.
  • Whitening alone captures most of the benefit, so even the cheaper two-step transform improves over the baseline.
  • The improvement persists across cluster counts from 500 to 2000, with larger gains at higher cluster counts.
  • Because ICA components are orthogonal and align with phonetic categories, the transformed space may support analysis of what information the speech model stores.
  • The dependence of the gains on the underlying self-supervised model means the preprocessing choice should be validated per model rather than assumed universal.

Reading between the lines

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

  • The paper does not test other discrete-unit consumers such as speech-to-speech translation or speech synthesis; if the same preprocessing yields better units, those pipelines could inherit the gains, but that is an extrapolation.
  • A direct way to test the anisotropy explanation is to measure cluster purity or mutual information between units and phones on raw versus whitened spaces; the paper's orthogonality histograms suggest such a check.
  • The bit-rate savings come partly from deduplication and BPE interacting with the unit stream; ablating those steps separately would clarify whether ICA's benefit is at the unit level or the compression level.
  • Since ICA with different number of components or different source priors, such as Student-t instead of Laplace, could change component geometry, tuning the ICA prior is a natural extension the paper leaves open.
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 / 4 minor

Summary. The paper studies four linear preprocessing transforms — standardization, PCA, whitening, and ICA — applied to self-supervised speech representations before k-means discretization into discrete speech units, evaluated on the Interspeech 2024 DSU challenge ASR track. Using XLS-R-300M features and a fixed E-Branchformer ASR baseline, it reports that whitening and ICA, combined with cosine-distance k-means, reduce CER and bit-rate relative to plain Euclidean k-means on LibriSpeech-100 and ML-SUPERB-1h. The paper also presents a cluster-count sweep and qualitative analyses of cluster centroid orthogonality and phonetic interpretability of ICA components.

Significance. The practical message — that a cheap, unsupervised linear preprocessing step can improve DSU quality and reduce bit-rate — would be useful to the speech-unit community. The paper uses the official DSU challenge benchmark, follows its evaluation protocol, and makes code available; the cluster-count sweep in Table II is a good robustness check. However, the central empirical result rests on single runs of a stochastic pipeline, and Table III shows the benefit is not universal across S3Ms. With repeated-seed evidence and a carefully scoped claim, this would be a solid empirical contribution.

major comments (4)
  1. [Section III-A, Tables I-III] All CER numbers in Tables I–III come from a single run of the full pipeline. Section III-A specifies 100 training epochs but no random seed, and the k-means preprocessing uses a randomly chosen 5% subset without a seed. The claimed advantages are 0.5–0.7 CER points on LibriSpeech subsets and 1.4–2.0 points on test 1h; these are within the typical seed-to-seed spread of end-to-end ASR training on this scale. Without repeated seeds, confidence intervals, or paired significance tests, the headline result that ICA+Cosine improves ASR is not distinguishable from run-to-run noise. Please report means and spread over at least three independent runs for the key Table I comparisons, ideally with paired differences where the same k-means units are used across ASR training seeds.
  2. [Abstract and Section V vs. Section III-D, Table III] The conclusion correctly restricts the claim to XLS-R-300M, but the abstract and introduction present the methods as generally effective (“demonstrate their effectiveness as preprocessing for k-means”). Table III shows ICA+Cosine degrades CER for mHuBERT (e.g., dev clean 5.6→7.6) and WavLM-Large (1.5→1.8), and also degrades test 1h for Wav2vec2.0-Large (27.2→30.4). The paper should state the scope of the claim in the abstract and provide at least a hypothesis-driven explanation or a diagnostic feature (e.g., anisotropy or singular-value spectrum) that predicts when ICA helps, rather than leaving “further exploration” entirely to future work.
  3. [Section IV-B, Figures 3-4] The qualitative interpretability analysis is selective: Figure 3 shows only centroids whose 10 nearest neighbors are all from the same phone, and Figure 4 shows only components whose top/bottom 5 extremes are from the same phone. With k=100 and 51 phones, random agreement could produce a substantial number of such matches, and no baseline or quantity is reported. The abstract’s claim that ICA components are interpretable needs a systematic measure (e.g., purity, phone-label agreement, or a permutation test) before it can be evaluated; as presented, these figures are anecdotal.
  4. [Section III-C, Table II] The stability claim that “ICA+Cosine outperforms Euclid by a consistent margin, except for 100 clusters” is based on single runs. At 100 clusters the comparison is mixed: ICA+Cosine is better on dev other and test other but worse on dev clean, test clean, and test 1h. Given the intended message of stability across cluster counts, this point should be qualified and ideally supported by repeated runs at the boundary value.
minor comments (4)
  1. [Section IV-A, Figure 2] The legend label “None (6.1×10 1)” is inconsistent with the text stating that similarities center around 0.6; presumably 6.1×10^{-1} is intended.
  2. [Section II-B, Eq. (6)] The notation “X(whiten) = S AΔT” and later “X(whiten) WΔΔT” would be clearer if the relationship between W and the mixing matrix A were stated explicitly, including the scaling convention.
  3. [Section IV-B] The description of the TIMIT analysis should specify which WavLM-Large layer was used and how phone-wise average pooling was performed; this is needed for reproducibility of the qualitative part.
  4. [Section I] Minor language issues include “S3Ms has become” in the first sentence and “The train and test set is the combination of these two corpora” in Section III-A; these should be corrected in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an empirical comparison of standard preprocessing methods against an external benchmark, with no derivation that reduces to its own inputs.

full rationale

The paper makes no formal derivation claim; its central claim is empirical: applying whitening or ICA before k-means improves discrete-unit ASR for XLS-R-300M on the DSU challenge benchmark. The ICA formulation (Eqs. 6-9) is standard theory cited from external sources (Cardoso 1997; Ono and Miyabe 2010), and the k-means/BPE/ASR pipeline follows the external DSU challenge baseline. The only author-group citations are to the benchmark and baseline system, which are independent external artifacts rather than assumptions equivalent to the conclusion. Hyperparameters such as the 5% k-means subsample are set uniformly across all compared methods, and the comparison is against an external baseline on standard test sets. The qualitative claims about ICA component interpretability are observations, not derived predictions. The absence of multiple seeds and significance tests is a statistical-evidence limitation, not a circularity, so it does not affect the circularity score.

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

The central empirical claim rests on hand-chosen hyperparameters (cluster count 2000, layer 17, ICA iterations 100, 5% training subset) and domain assumptions about the ICA mixing model, the Laplace source prior, and the validity of ASR CER/bit-rate as DSU quality proxies. None of these are derived in the paper, and no new entities are introduced.

free parameters (4)
  • number of k-means clusters = 2000
    Set following the DSU challenge baseline; Table II shows the ICA vs Euclid ranking reverses at 100 clusters, so the central result depends on this choice.
  • S3M layer for DSU extraction = 17th layer of XLS-R-300M
    Chosen based on top-3 DSU challenge systems; the paper does not search layers, and the ICA benefit may be layer-dependent.
  • ICA iteration count = 100
    Set by hand (Section III-A); no convergence analysis is reported.
  • k-means training subset = 5% of training set
    Footnote 3 states the authors empirically found 5% gives similar ASR to the baseline's 15%, a post-hoc choice that could affect the comparison.
assumptions (4)
  • domain assumption S3M representations approximately follow a linear mixing model with independent components (Eq. 6: X(whiten) = S A^T)
    Required for ICA preprocessing to yield meaningful independent axes; not verified for speech representations.
  • domain assumption The Laplace distribution for independent components (Eq. 8) matches the statistics of the whitened representations
    Standard ICA assumption; the quality of the demixing estimate depends on it.
  • domain assumption The DSU challenge ASR track (CER) and bit-rate (Eq. 10) are valid proxies for DSU quality
    Only one downstream task, ASR, is evaluated; conclusions may not transfer to synthesis, speech-to-speech translation, or textless NLP.
  • domain assumption Cosine-distance k-means centroids of TIMIT phone-wise pooled features reflect phonetic categories
    The interpretability analysis (Figs. 3-4) relies on nearest-neighbor agreement with TIMIT phone labels; no quantitative alignment metric is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discrete Speech Unit Extraction via Independent Component Analysis." pith.science (2026). https://pith.science/paper/VWU2QO5B

@misc{pith2026250106562,
  author       = {Pith},
  title        = {Pith review of: Discrete Speech Unit Extraction via Independent Component Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWU2QO5B}},
  note         = {Machine review of arXiv:2501.06562}
}
read the original abstract

Self-supervised speech models (S3Ms) have become a common tool for the speech processing community, leveraging representations for downstream tasks. Clustering S3M representations yields discrete speech units (DSUs), which serve as compact representations for speech signals. DSUs are typically obtained by k-means clustering. Using DSUs often leads to strong performance in various tasks, including automatic speech recognition (ASR). However, even with the high dimensionality and redundancy of S3M representations, preprocessing S3M representations for better clustering remains unexplored, even though it can affect the quality of DSUs. In this paper, we investigate the potential of linear preprocessing methods for extracting DSUs. We evaluate standardization, principal component analysis, whitening, and independent component analysis (ICA) on DSU-based ASR benchmarks and demonstrate their effectiveness as preprocessing for k-means. We also conduct extensive analyses of their behavior, such as orthogonality or interpretability of individual components of ICA.

Figures

Figures reproduced from arXiv: 2501.06562 by the authors.

Figure 1
Figure 1. Schematic illustration of processing flow for DSU-based ASR model. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparing cosine similarities and their averages between [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparing k-means centroids with phones with k = 100. We plot the centroids where the 10 nearest neighbor representations are from the same phone. Due to space limitations, we plot stops and fricatives only. methods in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 43 canonical work pages

  1. [1]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Proc. NeurIPS, 2020

    work page 2020

  2. [2]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W. Hsu, B. Bolte, Y . H. Tsai, et al. , “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE Trans. Audio, Speech, Lang. Process. , 2021

    work page 2021

  3. [3]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE J. Sel. Topics Signal Process., 2021

    work page 2021

  4. [4]

    XLS-R: Self-supervised cross- lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, et al. , “XLS-R: Self-supervised cross- lingual speech representation learning at scale,” in Proc. Interspeech , 2022

    work page 2022

  5. [5]

    SUPERB: Speech processing universal performance benchmark,

    S. Yang, P. Chi, Y . Chuang, et al. , “SUPERB: Speech processing universal performance benchmark,” in Proc. Interspeech, 2021

    work page 2021

  6. [6]

    SUPERB-SG: Enhanced speech processing universal performance benchmark for semantic and genera- tive capabilities,

    H. Tsai, H. Chang, W. Huang, et al. , “SUPERB-SG: Enhanced speech processing universal performance benchmark for semantic and genera- tive capabilities,” in Proc. ACL, 2022

    work page 2022

  7. [7]

    ML-SUPERB: Multilingual speech universal performance benchmark,

    J. Shi, D. Berrebbi, W. Chen, et al., “ML-SUPERB: Multilingual speech universal performance benchmark,” in Proc. Interspeech, 2023

    work page 2023

  8. [8]

    Dynamic-SUPERB: Towards a dy- namic, collaborative, and comprehensive instruction-tuning benchmark for speech,

    C. Huang, K. Lu, S. Wang, et al. , “Dynamic-SUPERB: Towards a dy- namic, collaborative, and comprehensive instruction-tuning benchmark for speech,” in Proc. ICASSP, 2024

    work page 2024

Show all 46 references
  1. [9]

    Layer-wise analysis of a self- supervised speech representation model,

    A. Pasad, J. Chou, and K. Livescu, “Layer-wise analysis of a self- supervised speech representation model,” in Proc. ASRU, 2021

    2021

  2. [10]

    Comparative layer-wise analysis of self-supervised speech models,

    A. Pasad, B. Shi, and K. Livescu, “Comparative layer-wise analysis of self-supervised speech models,” in Proc. ICASSP, 2023

    2023

  3. [11]

    Understanding probe behaviors through variational bounds of mutual information,

    K. Choi, J. Jung, and S. Watanabe, “Understanding probe behaviors through variational bounds of mutual information,” in Proc. ICASSP , 2024

    2024

  4. [12]

    Opening the black box of wav2vec feature encoder,

    K. Choi and E. Yeo, “Opening the black box of wav2vec feature encoder,” arXiv preprint arXiv:2210.15386 , 2022

    2022 arXiv

  5. [13]

    Self-supervised speech repre- sentations are more phonetic than semantic,

    K. Choi, A. Pasad, T. Nakamura, et al. , “Self-supervised speech repre- sentations are more phonetic than semantic,” in Proc. Interspeech, 2024

    2024

  6. [14]

    The Interspeech 2024 challenge on speech processing using discrete units,

    X. Chang, J. Shi, J. Tian, Y . Wu, Y . Tang, Y . Wu, S. Watanabe, Y . Adi, X. Chen, and Q. Jin, “The Interspeech 2024 challenge on speech processing using discrete units,” in Proc. Interspeech , pp. 2559–2563, 2024

    2024

  7. [15]

    Exploring speech recognition, translation, and understanding with discrete speech units: A comparative study,

    X. Chang, B. Yan, K. Choi, et al. , “Exploring speech recognition, translation, and understanding with discrete speech units: A comparative study,” in Proc. ICASSP, 2024

    2024

  8. [16]

    AudioLM: A language modeling approach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, et al. , “AudioLM: A language modeling approach to audio generation,” IEEE Trans. Audio, Speech, Lang. Process., 2023

    2023

  9. [17]

    Direct speech-to-speech translation with discrete units,

    A. Lee, P. Chen, C. Wang, et al. , “Direct speech-to-speech translation with discrete units,” in Proc. ACL, 2022

    2022

  10. [18]

    UnitY: Two-pass direct speech-to-speech translation with discrete units,

    H. Inaguma, S. Popuri, I. Kulikov, et al. , “UnitY: Two-pass direct speech-to-speech translation with discrete units,” in Proc. ACL, 2023

    2023

  11. [19]

    Au- dioPaLM: A large language model that can speak and listen,

    P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyen, et al. , “Au- dioPaLM: A large language model that can speak and listen,” arXiv preprint arXiv:2306.12925, 2023

    2023 arXiv

  12. [20]

    V oxtLM: Unified decoder-only models for consolidating speech recognition, synthesis and speech, text continuation tasks,

    S. Maiti, Y . Peng, S. Choi, et al. , “V oxtLM: Unified decoder-only models for consolidating speech recognition, synthesis and speech, text continuation tasks,” in Proc. ICASSP, 2024

    2024

  13. [21]

    On generative spoken language modeling from raw audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, et al. , “On generative spoken language modeling from raw audio,” Trans. ACL, 2021

    2021

  14. [22]

    Textually pretrained speech language models,

    M. Hassid, T. Remez, T. A. Nguyen, et al., “Textually pretrained speech language models,” in Proc. NeurIPS, 2024

    2024

  15. [23]

    A comparison of discrete and soft speech units for improved voice conversion,

    B. Van Niekerk, M.-A. Carbonneau, J. Za ¨ıdi, et al. , “A comparison of discrete and soft speech units for improved voice conversion,” in Proc. ICASSP, 2022

    2022

  16. [24]

    Improved distributed principal component analysis,

    Y . Liang, M.-F. Balcan, V . Kanchanapally, and D. Woodruff, “Improved distributed principal component analysis,” in Proc. NeurIPS, 2014

    2014

  17. [25]

    Analyzing acoustic word embeddings from pre-trained self-supervised speech models,

    R. Sanabria, H. Tang, and S. Goldwater, “Analyzing acoustic word embeddings from pre-trained self-supervised speech models,” in Proc. ICASSP, 2023

    2023

  18. [26]

    Wav2vec behind the scenes: How end2end models learn phonetics.,

    T. tom Dieck, P. A. P ´erez-Toro, T. Arias, et al. , “Wav2vec behind the scenes: How end2end models learn phonetics.,” in Proc. Interspeech , 2022

    2022

  19. [27]

    Orthogonality and isotropy of speaker and phonetic information in self-supervised speech representations,

    M. Mohamed, O. D. Liu, H. Tang, and S. Goldwater, “Orthogonality and isotropy of speaker and phonetic information in self-supervised speech representations,” in Proc. Interspeech, 2024

    2024

  20. [28]

    Discovering universal geometry in embeddings with ICA,

    H. Yamagiwa, M. Oyama, and H. Shimodaira, “Discovering universal geometry in embeddings with ICA,” in Proc. EMNLP, 2023

    2023

  21. [29]

    w2v-BERT: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,

    Y .-A. Chung, Y . Zhang, W. Han, et al. , “w2v-BERT: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,” in Proc. ASRU, pp. 244–250, 2021

    2021

  22. [30]

    Whitening sentence representations for better semantics and faster retrieval,

    J. Su, J. Cao, W. Liu, and Y . Ou, “Whitening sentence representations for better semantics and faster retrieval,” arXiv preprint arXiv:2103.15316 , 2021

    2021 arXiv

  23. [31]

    WhiteningBERT: An easy unsupervised sentence embedding approach,

    J. Huang, D. Tang, W. Zhong, et al. , “WhiteningBERT: An easy unsupervised sentence embedding approach,” in Proc. EMNLP, 2021

    2021

  24. [32]

    A review of blind source separation methods: Two two converging routes to ILRMA originating from ICA and NMF,

    H. Sawada, N. Ono, H. Kameoka, et al. , “A review of blind source separation methods: Two two converging routes to ILRMA originating from ICA and NMF,” APSIPA Trans. Signal Inf. Process., vol. 8, no. e12, pp. 1–14, 2019

    2019

  25. [33]

    Infomax and maximum likelihood for blind source separation,

    J.-F. Cardoso, “Infomax and maximum likelihood for blind source separation,” IEEE Signal Process. Lett. , vol. 4, no. 4, pp. 112–114, 1997

    1997

  26. [34]

    Auxiliary-function-based independent compo- nent analysis for super-gaussian sources,

    N. Ono and S. Miyabe, “Auxiliary-function-based independent compo- nent analysis for super-gaussian sources,” in Proc. LVA/ICA, pp. 165– 172, 2010

    2010

  27. [35]

    LibriSpeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “LibriSpeech: An ASR corpus based on public domain audio books,” in Proc. ICASSP , pp. 5206–5210, 2015

    2015

  28. [36]

    The Chi- naTelecom Interspeech2024 discrete speech unit asr challenge system description,

    Z. Li, Y . Yang, X. Li, J. Kang, X.-L. Zhang, and J. Li, “The Chi- naTelecom Interspeech2024 discrete speech unit asr challenge system description,” in Submission for Interspeech 2024 Challenge on Speech Processing Using Discrete Units , 2024

    2024

  29. [37]

    Exploration of efficient end-to-end ASR using discretized input from self-supervised learning,

    X. Chang, B. Yan, Y . Fujita, T. Maekaku, and S. Watanabe, “Exploration of efficient end-to-end ASR using discretized input from self-supervised learning,” in Proc. Interspeech, pp. 1399–1403, 2023

    2023

  30. [38]

    E-Branchformer: Branchformer with enhanced merging for speech recognition,

    K. Kim, F. Wu, Y . Peng, J. Pan, P. Sridhar, K. J. Han, and S. Watanabe, “E-Branchformer: Branchformer with enhanced merging for speech recognition,” pp. 84–91, 2023

    2023

  31. [39]

    Anisotropy is inherent to self-attention in transformers,

    N. Godey, ´E. V . de la Clergerie, and B. Sagot, “Anisotropy is inherent to self-attention in transformers,” 2024

    2024

  32. [40]

    Codec-superb@ slt 2024: A lightweight benchmark for neural audio codec models,

    H. Wu, X. Chen, Y .-C. Lin, et al. , “Codec-superb@ slt 2024: A lightweight benchmark for neural audio codec models,” in Proc. SLT , 2024

    2024

  33. [41]

    Textless speech-to- speech translation on real data,

    A. Lee, H. Gong, P.-A. Duquenne, H. Schwenk, P.-J. Chen, C. Wang, S. Popuri, Y . Adi, J. Pino, J. Gu, and W.-N. Hsu, “Textless speech-to- speech translation on real data,” in Proc. NAACL HLT , pp. 860–872, 2022

    2022

  34. [42]

    All-but-the-top: Simple and effective postpro- cessing for word representations,

    J. Mu and P. Viswanath, “All-but-the-top: Simple and effective postpro- cessing for word representations,” in Proc. ICLR, 2018

    2018

  35. [43]

    Effective dimensionality reduction for word embeddings,

    V . Raunak, V . Gupta, and F. Metze, “Effective dimensionality reduction for word embeddings,” in Proc. RepL4NLP@ACL, 2019

    2019

  36. [44]

    DARPA TIMIT: Acoustic- phonetic continuous speech corpus CD-ROM, NIST speech disc 1-1.1,

    J. Garofolo, L. Lamel, W. Fisher, et al. , “DARPA TIMIT: Acoustic- phonetic continuous speech corpus CD-ROM, NIST speech disc 1-1.1,” 1993

    1993

  37. [45]

    An information-theoretic analysis of self-supervised discrete representations of speech,

    B. M. Abdullah, M. M. Shaik, B. M ¨obius, and D. Klakow, “An information-theoretic analysis of self-supervised discrete representations of speech,” in Proc. Interspeech, 2023

    2023

  38. [46]

    Analysing discrete self supervised speech representation for spoken language modeling,

    A. Sicherman and Y . Adi, “Analysing discrete self supervised speech representation for spoken language modeling,” in Proc. ICASSP, 2023

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.