Pith. sign in

REVIEW 3 major objections 4 minor 39 references

ZeroSyl: Simple Zero-Resource Syllable Tokenization for Spoken Language Modeling

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read ZeroSyl claims that syllable boundaries for spoken language modeling can be detected with no training at all, simply by finding peaks in the L2 norm of frozen speech-model features, and that the resulting units beat prior trained syllabic t

desk verdict ZeroSyl is a refreshingly simple and mostly well-executed syllable tokenizer, but the headline LM comparisons are all on a development set with tuned hyperparameters, so the reported margins should be read as provisional. read the letter →

arxiv 2602.15537 v2 pith:OTH6HTFO submitted 2026-02-17 cs.CL eess.AS

Nicol Visser , Simon Malan , Danel Slabbert , Herman Kamper This is my paper · ORCID
classification cs.CLeess.AS
keywords spokenlanguagemodelingsyllabletokenizationzero-resourceWavLMself-supervisedspeechL2normpeakdetectionboundarymodelscaling
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 sets out to prove that syllable tokenization for pure speech language models does not require any training of the tokenizer. It claims that the L2 norm of framewise features from a single intermediate layer of a frozen self-supervised speech encoder, smoothed and scanned for prominent peaks, marks syllable boundaries as accurately as systems that are explicitly fine-tuned for the task. Mean-pooling a later layer's features inside those boundaries, then clustering with spherical K-means, yields a discrete syllabic vocabulary. Trained on top of these units, a standard causal language model outperforms prior syllabic tokenizers on lexical, syntactic, and narrative benchmarks, and the paper argues these syllable units have better scaling behaviour for syntax than fine-grained frame-level units. If correct, high-quality spoken language modeling is simpler than the field's recent multi-stage pipelines suggest.

What carries the argument

The load-bearing object is the L2-norm peak detector applied to layer 13 of WavLM, a frozen self-supervised speech encoder. The paper's key observation is that, unlike cosine distance between adjacent frames (the usual boundary signal), the norm of the frame itself pulses at syllable onset. This is combined with mean-pooling of layer-22 semantic features inside the detected segments, spherical K-means clustering with K-means++ initialization to form the vocabulary, and an unsupervised hierarchical-clustering step that collapses multiple silence centroids into one token. Together these steps replace a multi-stage fine-tuning pipeline with a single forward pass plus clustering.

What would settle it

Run the exact same pipeline on a held-out language or corpus and compare token-F1 against a cosine-distance peak detector; if the L2-norm signal does not exceed that baseline, or if shuffling the frame-wise L2 norms leaves the detected boundaries unchanged, the claim that the norm itself carries syllable structure is falsified.

Watch

Extended reading notes

Core claim

ZeroSyl's central claim is that syllable structure is already present, and readable, in the norms of a frozen self-supervised speech encoder's hidden features. For a given utterance, computing the L2 norm per frame at layer 13, smoothing with a window of 3, and taking peaks with prominence above 0.45σ produces syllable boundaries that reach an R-value of 75 and a token F1 of 54 on the evaluated data. The discovered segments are then represented by mean-pooled layer-22 features and clustered into a 10,000-centroid vocabulary (silence entries collapsed to one item), and a symbolic language model is trained on the resulting sequences. On the sWUGGY, sBLIMP, and tSC benchmarks, these units beat

Load-bearing premise

The paper assumes the three boundary-detection settings (layer 13, smoothing window 3, prominence threshold 0.45σ) tuned on development data give an unbiased measure of performance on the evaluation set, since it reports no held-out test set and no variance or error bars.

Editorial extensions

If this is right

  • Syllable-aware speech language models can be trained with no tokenizer fine-tuning: a frozen encoder, a norm, and clustering are sufficient.
  • The 52 bps bitrate means sequence lengths are around four times shorter than frame-level systems, easing long-range dependency modeling and lowering compute.
  • Silence collapsing, done unsupervised, adds a free boost to downstream LM scores and to per-syllable purity of clusters.
  • If the scaling result holds, syllable units are a better inductive bias than frame-level units for syntactic and narrative modeling as data grows.

Reading between the lines

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

  • The L2-norm pulse may reflect sonority or energy contours that are language-general; if so, the same detector could work on languages with different syllable structures without retuning—which the paper does not test.
  • ZeroSyl's units could be layered with unsupervised word discovery to build a hierarchy of word- and syllable-level tokens, potentially improving lexical tasks where syllable units currently lose to fine-grained tokens.
  • A concrete extension would be to vary the encoder layer and threshold per language or genre; because one layer and one threshold were chosen on development data, it remains open whether the signal is robust across domains.
  • Because the boundary signal requires no labels, ZeroSyl can be applied directly to low-resource languages, but performance there is untested in the paper.
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 / 4 minor

Summary. The paper proposes ZeroSyl, a training-free syllable tokenizer that detects boundaries by thresholding prominence peaks in the L2 norm of frozen WavLM Large layer-13 features, mean-pools layer-22 features between detected boundaries, and discretizes the pooled embeddings with spherical K-means (K=10,000, with a heuristic silence-token merge). An OPT-125M causal language model is then trained on the resulting token sequences. On LibriSpeech boundary detection, ZeroSyl reports R-value 75 and token F1 54, competitive with SyllableLM and better than Sylber. On spoken-language-modeling benchmarks with 6k hours of Libri-Light training, ZeroSyl reports the best scores among syllable systems on sWUGGY (68.0), sBLIMP (60.5), and tSC (68.1), with the lowest bitrate (52 bps). Scaling experiments show ZeroSyl trails the fine-grained SpidR on lexical tasks but has a steeper syntactic improvement with data scale. The central claim is that high-quality syllable discovery and competitive spoken language modeling can be achieved without complex multi-stage training.

Significance. If the reported results hold up, the contribution is valuable: it replaces multi-stage syllabic tokenizers with a simple, frozen-feature pipeline, and it demonstrates a new signal source (L2 norms of SSL hidden states) for boundary detection. The paper releases code and models, and the boundary detection evaluation is performed on LibriSpeech test sets. The method is unsupervised and does not use benchmark labels to set the pipeline, so the circularity burden is low. The main risk is that the headline LM comparisons are reported on the sLM21 development set, with several hyperparameters selected on that same set, so the margins over prior systems may not be robust. The scaling results, especially the sBLIMP trajectory, are interesting but need the same uncertainty quantification.

major comments (3)
  1. [§4.3, Table 2] The headline claim that ZeroSyl outperforms prior syllable tokenizers is supported entirely by LM scores on the sLM21 development set. The pipeline hyperparameters — WavLM boundary layer (13), smoothing window (3), prominence threshold (0.45σ), embedding layer (22), and the silence-collapse decision — were selected on that same set (Sections 3.1–3.2), and footnote 1 says the mean-per-token scoring convention was chosen because it improves sWUGGY and tSC on that set. With no held-out test split, confidence intervals, or significance tests, the reported margins (e.g., tSC 68.1 vs. 65.8) may reflect tuning rather than a genuine advantage. Please evaluate on the held-out sLM21 test set if available; otherwise report bootstrap intervals over utterances and/or a small number of LM training seeds, and show that the conclusions are stable across the hyperparameter grid considered.
  2. [§3.2] The silence-collapse step is load-bearing: Table 2 shows it improves sWUGGY from 67.0 to 68.0 and sBLIMP from 58.6 to 60.5. The decision to merge the smaller agglomerative branch is based on 'informal inspection' that this branch corresponds to silence. This is an untested assumption about cluster geometry. Please validate it quantitatively, e.g., by comparing the smaller-branch centroids against forced-alignment silence labels on LibriSpeech, and report how stable the branch assignment is across different K-means initializations, vocabulary sizes, and recording conditions.
  3. [§3.4, Table 2] The baseline selection procedure for SyllableLM is described as 'we consider all configurations (5.0, 6.25, and 8.33 Hz) and report the best result,' but only the 6.25 Hz row appears in Table 2. If 'best' is chosen per benchmark on the development set, the baseline is co-optimized on the same data as ZeroSyl, which should be stated explicitly; if 6.25 Hz is best for all benchmarks, the other configurations should be reported in a supplement. Without this, the reader cannot assess whether the comparison is apples-to-apples.
minor comments (4)
  1. [Table 2] The header contains a typo: 'V ocab. size' should be 'Vocab. size'. Also, 'ZeroSyl' vocabulary size is listed as '9k' while the text states 9,116; please be consistent.
  2. [Figure 3] The caption does not label the panels, while the text refers to 'top-left', 'top-right', and 'bottom-left'; there is no description of the bottom-right panel if it exists. Clarify the layout or redraw the figure.
  3. [Footnote 1] The capitalization is inconsistent ('Zerosyl-uncollapsed' vs. 'ZeroSyl (uncollapsed)'). More substantively, the unnormalized sBLIMP numbers are reported in the footnote but not discussed in the main text; if they change any of the cross-system rankings, they should be integrated into the results.
  4. [§4.1] The statement that 'boundaries that border silences are not evaluated' should be reconciled with the tokenization pipeline, where silence tokens are merged as a post-processing step. It would help to specify how this evaluation choice interacts with the silence-collapse heuristic.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ZeroSyl's boundaries and tokens are produced by unsupervised peak detection and clustering, and its benchmark comparisons are external to those fitting steps.

full rationale

The claimed derivation chain — L2-norm prominence peak detection on frozen WavLM layer 13 for boundaries, mean-pooling layer 22 features within boundaries, spherical K-means discretization, and OPT training — does not use any benchmark label or target score as an input. The boundary detector uses no supervision (Section 3.1); K-means is trained on unlabeled LibriSpeech audio (Section 3.2); silence collapsing is unsupervised; and language-model scores are computed on the external sLM21 development set after these components are frozen. Hyperparameter choices (layer 13, smoothing window 3, prominence 0.45σ, layer 22, silence collapse, mean-per-token scoring) are selected using development data, which is a statistical/overfitting concern about reported margins, not a circularity: the reported quantities are not equivalent by construction to any fitted parameter. Self-citations [9], [21], [22] are context or preliminary motivation, and the critical comparison to PromSeg is reproduced in Table 1 rather than relying on the self-citation. No step reduces to its own input, so no circularity is identified.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's central premise is that L2 norms of a frozen SSL model carry syllabic structure. The paper tunes several hyperparameters on development data and relies on an informally validated silence-merge heuristic. No new theoretical entities are introduced.

free parameters (5)
  • WavLM boundary layer = 13
    Selected on development data as best for R-value/token-F1 (Section 3.1).
  • Smoothing window size = 3
    Selected on development data to reduce noise (Section 3.1).
  • Prominence threshold δ = 0.45σ
    Selected on development data for peak detection (Section 3.1).
  • WavLM embedding layer = 22
    Chosen empirically: highest mutual information with syllable labels after clustering (Section 3.2).
  • Time-shift for boundary alignment = tuned per system
    Tuned per system on development data to account for representation latency (Section 4.1).
assumptions (5)
  • domain assumption WavLM Large (frozen) provides features whose L2 norms mark syllable boundaries and whose late-layer embeddings support K-means clustering into useful units.
    Core premise of the method; relies on a pretrained SSL model rather than deriving from first principles.
  • domain assumption Ground-truth syllables from Montreal Forced Aligner phone alignments plus the Gorman ARPABET syllabifier are accurate enough for evaluation.
    Used in Section 4.1 for boundary and token metrics; errors in these references propagate to reported scores.
  • ad hoc to paper The smaller branch in agglomerative hierarchical clustering of centroids corresponds to silence.
    Section 3.2: authors state this is known from informal inspection; no quantitative validation provided.
  • domain assumption sLM21 development set scores are representative of model quality.
    All LM benchmarks in Table 2 are on the development set, not a held-out test set (Section 4.3).
  • domain assumption LibriSpeech/Libri-Light English audiobook speech is sufficient to validate a universal 'zero-resource' tokenizer.
    K-means, LM training, and evaluation all use one language/corpus family; cross-lingual validity is not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZeroSyl: Simple Zero-Resource Syllable Tokenization for Spoken Language Modeling." pith.science (2026). https://pith.science/paper/OTH6HTFO

@misc{pith2026260215537,
  author       = {Pith},
  title        = {Pith review of: ZeroSyl: Simple Zero-Resource Syllable Tokenization for Spoken Language Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTH6HTFO}},
  note         = {Machine review of arXiv:2602.15537}
}
read the original abstract

Pure speech language models aim to learn language directly from raw audio without textual resources. A key challenge is that discrete tokens from self-supervised speech encoders result in excessively long sequences, motivating recent work on syllable-like units. However, methods like Sylber and SyllableLM rely on intricate multi-stage training pipelines. We propose ZeroSyl, a simple training-free method to extract syllable boundaries and embeddings directly from a frozen WavLM model. Using L2 norms of features in WavLM's intermediate layers, ZeroSyl achieves competitive syllable segmentation performance. The resulting segments are mean-pooled, discretized using K-means, and used to train a language model. ZeroSyl outperforms prior syllabic tokenizers across lexical, syntactic, and narrative benchmarks. Scaling experiments show that while finer-grained units are beneficial for lexical tasks, our discovered syllabic units exhibit better scaling behavior for syntactic modeling.

Figures

Figures reproduced from arXiv: 2602.15537 by the authors.

Figure 1
Figure 1. ZeroSyl detects syllabic boundaries using prominence￾based peak detection on features from layer 13 of a frozen WavLM. It then mean pools semantic features from layer 22 within the discovered boundaries and clusters these using spher￾ical K-means. A language model is trained on these cluster IDs. In this paper, we propose an unsupervised approach to syl￾lable tokenization that does not require the complex methods of… view at source ↗
Figure 2
Figure 2. Peak detection on the smoothed L2 norms of framewise features gives our predicted syllable boundaries. The dashed lines indicate the ground truth syllables. mean-pooling within each segment: ej = MeanPool(H [22] bj :bj+1 ) (1) We select this layer because higher WavLM layers are known to capture richer semantic information than those used for boundary detection [23]. Empirically, we find that layer 22 yields the hig… view at source ↗
Figure 3
Figure 3. In the sWUGGY task (Figure 3 top-left), ZeroSyl trails [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

39 extracted references · 5 linked inside Pith

  1. [1]

    Introduction Advances in self-supervised learning (SSL) [1] have made it possible to train language models directly on audio data: in- stead of using tokenized text, a causal language model is trained on tokens derived from an SSL model [2]. Without relying on any textual resources, this framework would enable natural lan- guage processing for languages l...

    2023

  2. [2]

    Standard benchmarks have been developed that assess how well these models capture lexical, syntactic, and semantic properties [15, 16]

    Related work This work focuses onpurespeech language models, where text is not used or incorporated at all [14]. Standard benchmarks have been developed that assess how well these models capture lexical, syntactic, and semantic properties [15, 16]. arXiv:2602.15537v1 [cs.CL] 17 Feb 2026 To date, AudioLM [4] achieves state-of-the-art performance in syntact...

    arXiv 2026

  3. [3]

    Models and methodology Our work builds on prominence-based segmentation [19]. How- ever, where previous work used the cosine similarity between adjacent frames to find word boundaries, we find that the L2 norm of the hidden representations themselves provides a signal corresponding to syllabic structure. In contrast to Sylber [10] and SyllableLM [11], our...

  4. [4]

    manufacturer,

    Experiments: Syllabic system comparison We perform a range of experiments to benchmark ZeroSyl’s units. In this section, we focus on systems that operate at the syllable level; comparisons against other systems are reserved for Section 5. We therefore compare ZeroSyl against Sylber [10] and SyllableLM [11] when trained on 6k hours of Libri-Light. This is ...

  5. [5]

    Experiments: Scaling behavior Finally, we investigate the performance of ZeroSyl as training data scales from 600 to 6k to 60k hours of Libri-Light. Here we compare mainly to SpidR [7], because it represents a well- performing non-syllabic approach which is still simple (like ZeroSyl): SpidR uses K-means to discretize features from a proposed SSL model. H...

  6. [6]

    By leveraging feature norms from a frozen SSL model (WavLM Large), ZeroSyl eliminates the complex multi-stage methods required by prior state-of-the-art syllabic tokenizers

    Conclusion In this work, we introduced ZeroSyl, a training-free framework for discovering syllabic units directly from unlabelled speech. By leveraging feature norms from a frozen SSL model (WavLM Large), ZeroSyl eliminates the complex multi-stage methods required by prior state-of-the-art syllabic tokenizers. ZeroSyl outperforms existing syllabic tokeniz...

  7. [7]

    Acknowledgements Nicol Visser and Simon Malan are funded through the Google PhD Fellowship Program

  8. [8]

    Self-Supervised Speech Representation Learning: A Review,

    A. Mohamed, H.-y. Lee, L. Borgholt, J. D. Havtorn, J. Edin, C. Igel, K. Kirchhoff, S.-W. Li, K. Livescu, L. Maaløe, T. N. Sainath, and S. Watanabe, “Self-Supervised Speech Representation Learning: A Review,”IEEE J. Sel. Top. Signal Process., vol. 16, no. 6, pp. 1179–1210, 2022

    2022

Show all 39 references
  1. [9]

    Self-Supervised Lan- guage Learning From Raw Audio: Lessons From the Zero Re- source Speech Challenge,

    E. Dunbar, N. Hamilakis, and E. Dupoux, “Self-Supervised Lan- guage Learning From Raw Audio: Lessons From the Zero Re- source Speech Challenge,”IEEE J. Sel. Top. Signal Process., vol. 16, no. 6, pp. 1211–1226, 2022

    2022

  2. [10]

    On Generative Spoken Language Modeling from Raw Audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, Y . Adi, A. Polyak, B. Bolte, T.-A. Nguyen, J. Copet, A. Baevski, A. Mohamed, and E. Dupoux, “On Generative Spoken Language Modeling from Raw Audio,” Trans. Assoc. Comput. Linguist., vol. 9, pp. 1336–1354, 2021

    2021

  3. [11]

    AudioLM: A Language Modeling Approach to Audio Generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi, and N. Zeghidour, “AudioLM: A Language Modeling Approach to Audio Generation,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 2523–2533, 2023

    2023

  4. [12]

    w2v-BERT: Combining Contrastive Learning and Masked Language Modeling for Self-Supervised Speech Pre-Training,

    Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu, “w2v-BERT: Combining Contrastive Learning and Masked Language Modeling for Self-Supervised Speech Pre-Training,” in Proc. ASRU, 2021, pp. 244–250

    2021

  5. [13]

    Scaling Properties of Speech Language Models,

    S. Cuervo and R. Marxer, “Scaling Properties of Speech Language Models,” inProc. EMNLP, 2024, pp. 351–361

    2024

  6. [14]

    SpidR: Learning Fast and Stable Linguistic Units for Spoken Language Models Without Supervision,

    M. Poli, M. Luthra, Y . Benchekroun, Y . Higuchi, M. Gleize, J. Shen, R. Algayres, Y .-A. Chung, M. Assran, J. Pino, and E. Dupoux, “SpidR: Learning Fast and Stable Linguistic Units for Spoken Language Models Without Supervision,”Trans. Mach. Learn. Res., 2025

    2025

  7. [15]

    Generative Spoken Language Model Based on Continuous Word-Sized Audio Tokens,

    R. Algayres, Y . Adi, T.-A. Nguyen, J. Copet, G. Synnaeve, B. Sagot, and E. Dupoux, “Generative Spoken Language Model Based on Continuous Word-Sized Audio Tokens,” inProc. EMNLP, 2023, pp. 3008–3028

    2023

  8. [16]

    Spoken Language Modeling with Duration-Penalized Self-Supervised Units,

    N. Visser and H. Kamper, “Spoken Language Modeling with Duration-Penalized Self-Supervised Units,” inProc. Interspeech, 2025, pp. 1968–1972

    2025

  9. [17]

    Sylber: Syllabic Embedding Repre- sentation of Speech from Raw Audio,

    C. J. Cho, N. Lee, A. Gupta, D. Agarwal, E. Chen, A. W. Black, and G. K. Anumanchipalli, “Sylber: Syllabic Embedding Repre- sentation of Speech from Raw Audio,” inProc. ICLR, 2025

    2025

  10. [18]

    SyllableLM: Learning Coarse Semantic Units for Speech Language Models,

    A. Baade, P. Peng, and D. Harwath, “SyllableLM: Learning Coarse Semantic Units for Speech Language Models,” inProc. ICLR, 2025

    2025

  11. [19]

    WavLM: Large-Scale Self- Supervised Pre-Training for Full Stack Speech Processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “WavLM: Large-Scale Self- Supervised Pre-Training for Full Stack Speech Processing,”IEEE J. Sel. Top. Signa...

    2022

  12. [20]

    OPT: Open Pre-trained Transformer Language Models,

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V . Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P. S. Koura, A. Sridhar, T. Wang, and L. Zettlemoyer, “OPT: Open Pre-trained Transformer Language Models,” 2022. [Online]....

    2022 arXiv

  13. [21]

    On The Landscape of Spoken Language Models: A Comprehensive Sur- vey,

    S. Arora, K.-W. Chang, C.-M. Chien, Y . Peng, H. Wu, Y . Adi, E. Dupoux, H.-y. Lee, K. Livescu, and S. Watanabe, “On The Landscape of Spoken Language Models: A Comprehensive Sur- vey,”Trans. Mach. Learn. Res., 2025

    2025

  14. [22]

    The Zero Resource Speech Challenge 2021: Spoken language modelling,

    E. Dunbar, M. Bernard, N. Hamilakis, T.-A. Nguyen, M. de Seyssel, P. Roz´e, M. Rivi`ere, E. Kharitonov, and E. Dupoux, “The Zero Resource Speech Challenge 2021: Spoken language modelling,” inProc. Interspeech, 2021, pp. 1574–1578

    2021

  15. [23]

    Textually Pretrained Speech Language Models,

    M. Hassid, T. Remez, T.-A. Nguyen, I. Gat, A. Conneau, F. Kreuk, J. Copet, A. Defossez, G. Synnaeve, E. Dupoux, R. Schwartz, and Y . Adi, “Textually Pretrained Speech Language Models,” inProc. NeurIPS, 2023, pp. 63 483–63 501

    2023

  16. [24]

    SD-HuBERT: Sentence-Level Self-Distillation In- duces Syllabic Organization in HuBERT,

    C. J. Cho, A. Mohamed, S.-W. Li, A. W. Black, and G. K. Anu- manchipalli, “SD-HuBERT: Sentence-Level Self-Distillation In- duces Syllabic Organization in HuBERT,” inProc. ICASSP, 2024, pp. 12 076–12 080

    2024

  17. [25]

    HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,

    W. Hsu, B. Bolte, Y . H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 29, pp. 3451–3460, 2021

    2021

  18. [26]

    What Do Self- Supervised Speech Models Know About Words?

    A. Pasad, C.-M. Chien, S. Settle, and K. Livescu, “What Do Self- Supervised Speech Models Know About Words?”Trans. Assoc. Comput. Linguist., vol. 12, pp. 372–391, 2024

    2024

  19. [27]

    A Computational Model for Unsuper- vised Word Discovery,

    L. ten Bosch and B. Cranen, “A Computational Model for Unsuper- vised Word Discovery,” inProc. Interspeech, 2007, pp. 1481–1484

    2007

  20. [28]

    Unsupervised Word Discovery: Boundary Detection with Clustering vs. Dynamic Pro- gramming,

    S. Malan, B. van Niekerk, and H. Kamper, “Unsupervised Word Discovery: Boundary Detection with Clustering vs. Dynamic Pro- gramming,” inProc. ICASSP, 2025

    2025

  21. [29]

    Should Top-Down Clustering Affect Boundaries in Unsupervised Word Discovery?

    ——, “Should Top-Down Clustering Affect Boundaries in Unsupervised Word Discovery?” 2025. [Online]. Available: https://arxiv.org/abs/2507.19204

    2025 arXiv

  22. [30]

    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,” inProc. ICASSP, 2023, pp. 1–5

    2023

  23. [31]

    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,” inProc. ICASSP, 2015, pp. 5206–5210

    2015

  24. [32]

    The Faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazar´e, M. Lomeli, L. Hosseini, and H. J´egou, “The Faiss library,”

  25. [34]

    Libri-Light: A Benchmark for ASR with Limited or No Supervi- sion,

    J. Kahn, M. Rivi`ere, W. Zheng, E. Kharitonov, Q. Xu, P. E. Mazar´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen, T. Likhoma- nenko, G. Synnaeve, A. Joulin, A. Mohamed, and E. Dupoux, “Libri-Light: A Benchmark for ASR with Limited or No Supervi- sion,” inProc. ICASSP...

    2020

  26. [35]

    Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Sondereg- ger, “Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,” inProc. Interspeech, 2017, pp. 498–502

    2017

  27. [36]

    Python module for syllabifying English ARPABET transcriptions,

    K. Gorman, “Python module for syllabifying English ARPABET transcriptions,” 2013, GitHub repository. [Online]. Available: https://github.com/kylebgorman/syllabify

    2013

  28. [37]

    An improved speech segmentation quality measure: the R-value,

    O. J. R¨as¨anen, U. K. Laine, and T. Altosaar, “An improved speech segmentation quality measure: the R-value,” inProc. Interspeech, 2009, pp. 1851–1854

    2009

  29. [38]

    On the robust automatic segmentation of spontaneous speech,

    B. Petek, O. Andersen, and P. Dalsgaard, “On the robust automatic segmentation of spontaneous speech,” inProc. ICSLP, 1996, pp. 913–916

    1996

  30. [39]

    data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,

    A. Baevski, W.-N. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,” 2022. [Online]. Available: https://arxiv.org/abs/2202.03555

    2022 arXiv

  31. [2024]

    Available: https://arxiv.org/abs/2401.08281

    [Online]. Available: https://arxiv.org/abs/2401.08281

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.