Pith. sign in

REVIEW 3 major objections 6 minor 46 references

findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding

T0 review · 3 major / 6 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Syllable speech tokenization is a design space: recombining classical envelopes and neural features beats fixed pipelines and works across languages, including newly annotated Kono.

desk verdict Solid toolkit paper: modular recombination of known syllabifiers is real and measured; annotation quality and missing artifact link are the only soft spots that matter. read the letter →

arxiv 2603.26292 v2 pith:ZUUGOSD5 submitted 2026-03-27 cs.CL cs.AI

Héctor Javier Vázquez Martínez This is my paper
classification cs.CLcs.AI
keywords syllablesegmentationspeechtokenizationrepresentationsbenchmarkingcross-lingualprocessingmodulartoolkitembeddings
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

Syllable-sized units cut speech sequences to a few tokens per second while keeping linguistic structure, which helps spoken language models and unsupervised word discovery. Research on how to find those units has been scattered across separate scripts, models, and evaluation recipes, so methods are hard to compare or reuse. This paper introduces findsylls, an open modular toolkit that puts classical amplitude-envelope detectors and modern self-supervised syllabifiers under one interface for segmentation, syllable embeddings, and multi-level scoring. The toolkit deliberately lets researchers swap envelopes, feature extractors, and segmentation algorithms, then measures nuclei, boundaries, and full spans on English, Spanish, and hand-annotated Kono. The central result is that recombining components can improve boundary and span accuracy over the published default pipelines, and that the same framework supports both high-resource and under-resourced settings.

What carries the argument

A three-module pipeline—envelope computation, frame-level feature extraction, and interchangeable segmentation algorithms—plus export of neural traces as pseudo-envelopes so any cue can drive the same peak detector or interval algorithm, evaluated with nuclei/boundary/span F1, token rate, and throughput.

What would settle it

Re-annotate a held-out sample of the same corpora (especially child-directed speech and Kono) with independent syllable intervals under a stated phonological theory, then re-run the Table 2 configurations: if the recombination gains (e.g., Sylber cosine cue + peakdetect over the published Sylber pipeline) disappear or reverse, the central claim fails.

Watch

Extended reading notes

Core claim

findsylls shows that syllabic tokenization is a design space rather than a fixed method: classical envelope detectors and neural syllabifiers can be standardized and recombined under one interface, and those recombinations produce measurable gains in boundary and span recovery over packaged defaults while remaining usable from English and Spanish down to newly annotated fieldwork data in Kono.

Load-bearing premise

The paper treats dictionary-based, forced-alignment, and hand-drawn syllable intervals as reliable ground truth under a fixed 50 ms tolerance; if those labels systematically disagree with the phonetic notion of syllable the methods target, the accuracy rankings and claimed recombination gains become unreliable.

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 introduces findsylls, an open-source modular toolkit that unifies classical amplitude-envelope syllable detectors and representation-driven syllabifiers (Sylber, VG-HuBERT, HuBERT, etc.) under a common interface for envelope computation, feature extraction, segmentation, syllable-level embedding, and multi-granular evaluation (nuclei, boundaries, spans). Components can be recombined (e.g., SSL-derived pseudo-envelopes with peakdetect; Sylber cues with alternative segmenters). Benchmarks cover English and Spanish adult and child-directed speech plus newly hand-annotated Kono fieldwork data. Table 2 and Figure 1 report that recombination can improve over published defaults (notably Sylber cosine cue + peakdetect raising boundary F1 from 63.1 to 69.9), that nuclei are much easier than spans, and that envelope baselines offer high RTFx while neural methods trade throughput for boundary/span quality.

Significance. If the toolkit is released and maintained as described, this is useful shared infrastructure for a fragmented subfield: matched metrics, modular ablations of representation vs. segmenter, and a path for low-resource work. Concrete strengths include (i) controlled component-swap results with reported F1 deltas rather than only packaged pipelines, (ii) multi-granular evaluation that exposes nuclei–boundary–span degradation, (iii) token-rate and relative RTFx reporting under a fixed setup, and (iv) new hand-aligned Kono material. The design-space framing of syllabic tokenization is well supported by the recombination experiments and is of practical interest for spoken language modeling and unsupervised word discovery.

major comments (3)
  1. Table 2 aggregates syllable-weighted Prec/Rec/F1 over seven corpora whose sizes differ by orders of magnitude (LS-100h: 1.37M syllables; Kono: 636; Brent: 1.5k). The reported “averaged over seven corpora” numbers are therefore dominated by LibriSpeech (and secondarily WikiSpanish). This weakens the load-bearing cross-lingual / under-resourced part of the claim in the abstract and §1/§6. Please add unweighted or per-corpus primary tables (or corpus-size-normalized averages) for nuclei/boundaries/spans, and state explicitly how much of Table 2 is driven by LS-100h. Figure 1a helps for boundaries but does not fully replace this for the main quantitative claims.
  2. §4.1–4.2: English syllable intervals are largely derived from a syllabified CMU dictionary (plus MFA phone alignments); Spanish from faseAlign; only Kono is fully hand-aligned. Span F1 remains low across all systems (~13–47). Because span and boundary rankings are central to the recombination story (e.g., Sylber+peakdetect best boundary/span in Table 2), the paper should quantify sensitivity of those rankings to annotation source or tolerance (beyond the planned error audit in §5.3)—e.g., a small hand-checked English/Spanish subset, or results under 30/50/70 ms tolerances. Absolute F1 need not be phonologically perfect for the modularity claim, but the paper currently over-reads aggregate F1 as method quality without enough annotation-robustness evidence.
  3. §4.1 and Table 2 RTFx column: classical baselines use CPU parallelization while neural configs run sequentially on Apple M1 Max / MPS. Relative ordering is informative, but the large RTFx gap (e.g., SBS 684× vs. MinCut 6×) mixes algorithmic cost with implementation and device choices. For the accuracy–throughput trade-off claimed in §5.2/§6, either (a) report CPU-only and GPU-only columns, or (b) clearly mark RTFx as non-comparable across classical vs. neural families and restrict throughput claims to within-family comparisons.
minor comments (6)
  1. §5.1: The poor VG-HuBERT CLS+threshold result (41.9 nuclei F1) is reported as an “observed baseline” without a short diagnostic (layer choice, threshold calibration, attention aggregation). A brief ablation or note that the configuration was not retuned would help readers avoid over-interpreting the failure.
  2. Table 1 / §4.2: Ornat-Swingley is evaluated for nuclei only; this should be flagged in Table 2’s caption and in any nuclei-only vs. full-suite averages so readers do not assume identical evaluation support across all seven corpora.
  3. Figure 1 caption and panels (b)–(c): Define ΔF1 polarity and the exact baseline of each swap in the caption (not only in the main text) so the figure is self-contained.
  4. §2–3: Several methods are described as “to come” or not yet included (§5.3). A short table of currently implemented vs. planned modules would set expectations for the v1 release.
  5. Presentation: normalize toolkit name formatting (findsylls vs. find-sylls vs. code font); fix minor spacing/typos (e.g., “findsyllsorganizes”, “W A V–TextGrid”); ensure all GitHub/data URLs and the syllabify/faseAlign footnotes remain stable in the camera-ready version.
  6. §4.1: State the exact peakdetect lookahead/auto-calibration rule and cosine/CLS thresholds used, or point to a config file in the release, so the Table 2 numbers are reproducible without reverse-engineering defaults.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical toolkit paper with measured recombination gains against external annotations, not definitional or fitted predictions.

full rationale

findsylls is a systems/tool paper whose central claims are (i) a modular interface that recombines published envelope, feature, and segmentation components and (ii) empirical multi-granular F1, token-rate, and RTFx comparisons under a fixed 50 ms protocol on English, Spanish, and hand-annotated Kono data. Table 2 and Figure 1b–c report measured deltas (e.g., Sylber cosine cue + peakdetect raising boundary F1 from 63.1 to 69.9) obtained by applying the same external TextGrid labels to every configuration; nothing is algebraically forced by construction, fitted to the evaluation targets, or justified solely by a self-citation uniqueness theorem. The single author cites prior external work (Sylber, VG-HuBERT, MinCut, peakdetect, etc.) as implementable baselines, not as load-bearing self-proofs. Annotation quality and default hyperparameters are acknowledged limitations (§5.3) but do not create circular reduction of the reported gains. Score 0 is therefore appropriate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

This is an empirical systems paper. Load-bearing content is not a formal derivation but a set of evaluation conventions (tolerance, ground-truth construction, default hyperparameters) plus the operational claim that classical and SSL components can be freely recombined under one API. Free parameters are the matching and segmentation knobs that directly move reported F1 and token rates. Domain assumptions concern what counts as a syllable label and that SSL layers carry usable syllabic structure. No new physical entities are postulated; the toolkit and pseudo-envelope constructions are engineering objects.

free parameters (4)
  • boundary matching tolerance = 50 ms
    Fixed at 50 ms for all nuclei/boundary/span scoring; chosen to follow prior work but directly controls true-positive counts and therefore all F1 numbers in Table 2.
  • MinCut expected syllable duration = 220 ms
    Recommended default of 220 ms used for all MinCut configurations; changes token rate and boundary placement.
  • peakdetect lookahead / thresholds = auto-calibrated defaults
    Auto-calibrated lookahead and Billauer-style peak criteria applied uniformly to classical and pseudo-envelopes; alters nuclei detection operating point.
  • cosine / CLS segmentation thresholds = published defaults
    Thresholds inherited from Sylber and VG-HuBERT pipelines for greedy merge and CLS-attention segmentation; not re-tuned per corpus in the main table.
assumptions (4)
  • domain assumption A fixed 50 ms temporal tolerance is an adequate matching criterion for syllable nuclei, boundaries, and spans across languages and recording conditions.
    Stated in §4.1 evaluation setup; all multi-granular F1 scores rest on it.
  • domain assumption Rule-based syllabification of forced-aligned phones (CMU+syllabify, MFA, faseAlign) yields valid syllable interval ground truth for English and Spanish adult and child-directed speech.
    §4.2 Corpora and Annotations; used for six of seven corpora.
  • domain assumption Self-supervised representations (HuBERT, VG-HuBERT, Sylber) encode syllabic structure that can be recovered by the listed segmentation algorithms without language-specific retraining.
    Background assumption imported from cited SSL syllabification papers and used for all neural rows in Table 2.
  • ad hoc to paper Default hyperparameters of the original methods, run under one fixed hardware/software stack, permit fair relative comparison of accuracy and RTFx.
    §4.1 experimental setup; limitations section notes implementation sensitivity.
invented entities (2)
  • findsylls modular envelope/feature/segmenter interface
    purpose: Unify classical and neural syllabifiers so components can be recombined and evaluated under one API.
    Core software contribution of the paper; independent evidence is limited to the reported benchmarks rather than an external physical prediction.
  • SSL-derived pseudo-envelopes (framewise cosine similarity, CLS-attention traces, featSSM row-mean coherence) independent evidence
    purpose: Allow the same peakdetect procedure used on classical amplitude envelopes to operate on neural cues for controlled comparison.
    Operational construction defined in §2.3 and §4.1; evaluated empirically but not a new physical object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding." pith.science (2026). https://pith.science/paper/ZUUGOSD5

@misc{pith2026260326292,
  author       = {Pith},
  title        = {Pith review of: findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUUGOSD5}},
  note         = {Machine review of arXiv:2603.26292}
}
read the original abstract

Syllable-level units offer compact and linguistically meaningful representations for spoken language modeling and unsupervised word discovery, but research on syllabification remains fragmented across disparate implementations, datasets, and evaluation protocols. We introduce findsylls, a modular, language-agnostic toolkit that unifies classical syllable detectors and end-to-end syllabifiers under a common interface for syllable segmentation, embedding extraction, and multi-granular evaluation. The toolkit implements and standardizes widely used methods (e.g., Sylber, VG-HuBERT) and allows their components to be recombined, enabling controlled comparisons of representations, algorithms, and token rates. We demonstrate findsylls on English and Spanish corpora and on new hand-annotated data from Kono, an underdocumented Central Mande language, illustrating how a single framework can support reproducible syllable-level experiments across both high-resource and under-resourced settings.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 2 canonical work pages

  1. [1]

    syllable

    Introduction Syllables provide a natural temporal unit for modeling speech, sitting between short analysis frames and longer lexical units such as words. Syllable-level representations have recently been shown to support efficient spoken language modeling, self- supervised word discovery, and downstream speech tasks, by drastically reducing sequence lengt...

    arXiv 2026

  2. [2]

    Methods findsyllsorganizes syllable processing into three inter- operable modules: (i) envelope computation, (ii) frame-level feature extraction, and (iii) segmentation algorithms. To- gether, these modules cover both classical signal-processing approaches and recent representation-driven methods, while allowing their components to be recombined within a ...

  3. [3]

    Sylber-style

    Thefindsyllstoolkit A key design goal offindsyllsis that these algorithms are not tied to any specific feature extractor. The greedy co- sine strategy originally used in Sylber can be applied to Hu- BERT or VG-HuBERT features, enabling direct tests of how “Sylber-style” segmentation behaves with alternative encoders. Similarly, the MinCut algorithm optimi...

  4. [4]

    Experiments 4.1. Experimental setup We compare two classical envelope-based baselines (SBS, Theta) with three widely used neural configurations (Syl- ber; VG-HuBERT featSSM+MinCut; VG-HuBERT CLS+CLS- threshold), and selected modular recombinations enabled by findsylls. To enable interoperability with envelope-based segmentation methods, we export SSL-deri...

  5. [5]

    Segmentation accuracy Table 2 reports syllable-weighted Precision, Recall, and F1 for syllable nuclei, boundaries, and spans aggregated across seven corpora

    Results and Discussion 5.1. Segmentation accuracy Table 2 reports syllable-weighted Precision, Recall, and F1 for syllable nuclei, boundaries, and spans aggregated across seven corpora. Figure 1a (left) shows boundary F1 broken down by dataset for the default baselines/configurations. Across config- urations, nuclei detection is substantially easier than ...

  6. [6]

    We demonstrated its capabilities by benchmarking ten classical and representation- driven configurations across seven corpora under matched met- rics

    Conclusion We introducedfindsylls(v1), a language-agnostic, mod- ular toolkit for syllable-level speech tokenization and embed- ding, with a unified interface for nuclei, boundary, and span seg- mentation and multi-granular evaluation. We demonstrated its capabilities by benchmarking ten classical and representation- driven configurations across seven cor...

  7. [7]

    All content, including technical descriptions, ex- perimental design, results interpretation, and conclusions, is original work of the authors

    Generative AI Use Disclosure Generative AI was used only for polishing and editing the manuscript. All content, including technical descriptions, ex- perimental design, results interpretation, and conclusions, is original work of the authors. The authors assume full responsi- bility for the accuracy and integrity of the paper and consent to its submission

  8. [8]

    Sylber: Syllabic Embedding Representation of Speech from Raw Audio,

    C. J. Cho, N. Lee, A. Gupta, D. Agarwal, E. Chen, A. Black, and G. Anumanchipalli, “Sylber: Syllabic Embedding Representation of Speech from Raw Audio,” Oct. 2024. [Online]. Available: https://openreview.net/forum?id=FyMjfDQ9RO

    2024

Show all 46 references
  1. [9]

    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,” Oct

  2. [10]

    Available: https://openreview.net/forum?id= dGSOn7sdWg

    [Online]. Available: https://openreview.net/forum?id= dGSOn7sdWg

  3. [11]

    Towards Unsupervised Speech Recognition at the Syllable-Level,

    L. Wang, J. Ni, K.-W. Chang, S. Bhati, D. Harwath, M. Hasegawa- Johnson, and J. R. Glass, “Towards Unsupervised Speech Recognition at the Syllable-Level,” Oct. 2025, arXiv:2510.03639 [cs] version: 1. [Online]. Available: http://arxiv.org/abs/2510. 03639

    2025

  4. [12]

    Scaling Spoken Language Models with Syllabic Speech Tokenization,

    N. Lee, C. J. Cho, A. W. Black, and G. K. Anumanchipalli, “Scaling Spoken Language Models with Syllabic Speech Tokenization,” Feb. 2026, arXiv:2509.26634 [cs]. [Online]. Available: http://arxiv.org/abs/2509.26634

    2026

  5. [13]

    Word Discovery in Visually Grounded, Self-Supervised Speech Models,

    P. Peng and D. Harwath, “Word Discovery in Visually Grounded, Self-Supervised Speech Models,” inInterspeech 2022. ISCA, Sep. 2022, pp. 2823–2827. [Online]. Available: https://www. isca-archive.org/interspeech 2022/peng22c interspeech.html

    2022

  6. [14]

    Syllable Discovery and Cross-Lingual Generalization in a Visually Grounded, Self-Supervised Speech Model,

    P. Peng, S.-W. Li, O. R ¨as¨anen, A. Mohamed, and D. Harwath, “Syllable Discovery and Cross-Lingual Generalization in a Visually Grounded, Self-Supervised Speech Model,” 2023, pp. 391–395. [Online]. Available: https://www.isca-archive.org/ interspeech 2023/peng23e interspeech.html

    2023

  7. [15]

    SD-HuBERT: Sentence-Level Self-Distillation Induces Syllabic Organization in Hubert,

    C. J. Cho, A. Mohamed, S.-W. Li, A. W. Black, and G. K. Anumanchipalli, “SD-HuBERT: Sentence-Level Self-Distillation Induces Syllabic Organization in Hubert,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Apr. 2024, pp....

    2024

  8. [16]

    Syllable rhythm in English and Mandarin,

    M. Liberman, “Syllable rhythm in English and Mandarin,” Feb

  9. [17]

    Available: https://languagelog.ldc.upenn.edu/nll/ ?p=58120

    [Online]. Available: https://languagelog.ldc.upenn.edu/nll/ ?p=58120

  10. [18]

    Unsupervised lexicon learning from speech is limited by representations rather than clustering,

    D. Adendorff, S. Malan, and H. Kamper, “Unsupervised lexicon learning from speech is limited by representations rather than clustering,” Oct. 2025, arXiv:2510.09225 [eess] version: 1. [Online]. Available: http://arxiv.org/abs/2510.09225

    2025

  11. [19]

    Automatic segmentation of speech into syllabic units,

    P. Mermelstein, “Automatic segmentation of speech into syllabic units,”The Journal of the Acoustical Society of America, vol. 58, no. 4, pp. 880–883, Oct. 1975

    1975

  12. [20]

    Robust acoustic-based syllable detection,

    Z. Xie and P. Niyogi, “Robust acoustic-based syllable detection,” inInterspeech 2006. ISCA: ISCA, Sep. 2006, pp. paper 1327– Wed1BuP.6–0. [Online]. Available: https://www.isca-archive. org/interspeech 2006/xie06 interspeech.html

    2006

  13. [21]

    Robust Speech Rate Estimation for Spontaneous Speech,

    D. Wang and S. S. Narayanan, “Robust Speech Rate Estimation for Spontaneous Speech,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 8, pp. 2190–2201, Nov

  14. [22]

    Available: https://ieeexplore.ieee.org/document/ 4317582

    [Online]. Available: https://ieeexplore.ieee.org/document/ 4317582

  15. [23]

    Robust speaking rate estimation using broad phonetic class recognition,

    J. Yuan and M. Liberman, “Robust speaking rate estimation using broad phonetic class recognition,” in2010 IEEE International Conference on Acoustics, Speech and Signal Processing, Mar. 2010, pp. 4222–4225. [Online]. Available: https://ieeexplore. ieee.org/document/5495686/authors

    2010

  16. [24]

    Automatic Syllable Segmen- tation Using Broad Phonetic Class Information,

    B. Ludusan and E. Dupoux, “Automatic Syllable Segmen- tation Using Broad Phonetic Class Information,”Procedia Computer Science, vol. 81, pp. 101–106, Jan. 2016. [On- line]. Available: https://www.sciencedirect.com/science/article/ pii/S1877050916300503

    2016

  17. [25]

    Speech rhythm guided syllable nuclei detection,

    Y . Zhang and J. R. Glass, “Speech rhythm guided syllable nuclei detection,” in2009 IEEE International Conference on Acoustics, Speech and Signal Processing, Apr. 2009, pp. 3797–

    2009

  18. [26]

    Available: https://ieeexplore.ieee.org/document/ 4960454

    [Online]. Available: https://ieeexplore.ieee.org/document/ 4960454

  19. [27]

    Syllables,

    M. Liberman, “Syllables,” Feb. 2022. [Online]. Available: https://languagelog.ldc.upenn.edu/nll/?p=46144

    2022

  20. [28]

    Pre-linguistic segmentation of speech into syllable-like units,

    O. R ¨as¨anen, G. Doyle, and M. C. Frank, “Pre-linguistic segmentation of speech into syllable-like units,”Cognition, vol. 171, pp. 130–150, Feb. 2018. [Online]. Available: https://www. sciencedirect.com/science/article/pii/S0010027717302901

    2018

  21. [29]

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

    W.-N. 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,”arXiv.org, Jun. 2021. [Online]. Available: https://arxiv.org/abs/2106.07447v1

    2021 arXiv

  22. [30]

    peakdet: Peak detection using MATLAB (non-derivative local extremum, maximum, minimum),

    E. Billauer, “peakdet: Peak detection using MATLAB (non-derivative local extremum, maximum, minimum),” Jan

  23. [31]

    Available: https://billauer.co.il/blog/2009/01/ peakdet-matlab-octave/

    [Online]. Available: https://billauer.co.il/blog/2009/01/ peakdet-matlab-octave/

    2009

  24. [32]

    Syll-O-Matic: An adaptive time-frequency representation for the automatic segmentation of speech into syllables,

    N. Obin, F. Lamare, and A. Roebel, “Syll-O-Matic: An adaptive time-frequency representation for the automatic segmentation of speech into syllables,” in2013 IEEE International Conference on Acoustics, Speech and Signal Processing, May 2013, pp. 6699–

    2013

  25. [33]

    Available: https://ieeexplore.ieee.org/document/ 6638958

    [Online]. Available: https://ieeexplore.ieee.org/document/ 6638958

  26. [34]

    Automatic blind syllable segmentation for continuous speech,

    R. Villing, J. Timoney, T. Ward, and J. Costello, “Automatic blind syllable segmentation for continuous speech,” inIrish Signals and Systems Conference 2004, vol. 2004. Belfast, Ireland: IEE, 2004, pp. 41–46. [Online]. Available: https://digital-library. theiet.org/content/con...

    2004 doi

  27. [35]

    TIMIT Acoustic- Phonetic Continuous Speech Corpus,

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, D. S. Pallett, N. L. Dahlgren, V . Zue, and J. G. Fiscus, “TIMIT Acoustic- Phonetic Continuous Speech Corpus,” 1993. [Online]. Available: https://catalog.ldc.upenn.edu/LDC93S1

    1993

  28. [36]

    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,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Apr. 2015, pp. 5206–

    2015

  29. [37]

    Available: https://ieeexplore.ieee.org/document/ 7178964

    [Online]. Available: https://ieeexplore.ieee.org/document/ 7178964

  30. [38]

    Wikipedia Spanish Speech and Transcripts,

    C. Hern ´andez Mena, “Wikipedia Spanish Speech and Transcripts,” Aug. 2021. [Online]. Available: https: //catalog.ldc.upenn.edu/LDC2021S07

    2021

  31. [39]

    The role of exposure to isolated words in early vocabulary development,

    M. R. Brent and J. M. Siskind, “The role of exposure to isolated words in early vocabulary development,”Cognition, vol. 81, no. 2, pp. B33–B44, Sep. 2001. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S0010027701001226

    2001

  32. [40]

    Segmenting words from natural speech: subsegmental variation in segmental cues,

    C. A. Rytting, C. Brew, and E. Fosler-Lussier, “Segmenting words from natural speech: subsegmental variation in segmental cues,”Journal of Child Language, vol. 37, no. 3, pp. 513–543, Jun. 2010. [Online]. Available: https://www.cambridge.org/core/ product/identifier/S030500091...

    2010

  33. [41]

    Philadelphia home corpus,

    D. Swingley, “Philadelphia home corpus,” Databrary, V olume 1801, 2024, retrieved March 2, 2026 from https://databrary.org/ volume/1801

    2024

  34. [42]

    L ´opez Ornat,La adquisici ´on de la lengua espa˜nola, 1st ed., ser

    S. L ´opez Ornat,La adquisici ´on de la lengua espa˜nola, 1st ed., ser. Ling¨u´ıstica y teor´ıa literaria. Madrid: Siglo Veintiuno de Espa ˜na Editores, 1994

    1994

  35. [43]

    Lexical Learning May Contribute to Phonetic Learning in Infants: A Cor- pus Analysis of Maternal Spanish,

    D. Swingley and C. Alarcon, “Lexical Learning May Contribute to Phonetic Learning in Infants: A Cor- pus Analysis of Maternal Spanish,”Cognitive Sci- ence, vol. 42, no. 5, pp. 1618–1641, 2018, eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/cogs.12620. [On- line]. Avai...

    2018 doi

  36. [44]

    Speech Model Pre-training for End-to-End Spoken Language Understanding,

    L. Lugosch, M. Ravanelli, P. Ignoto, V . S. Tomar, and Y . Bengio, “Speech Model Pre-training for End-to-End Spoken Language Understanding,” Jul. 2019, arXiv:1904.03670 [eess]. [Online]. Available: http://arxiv.org/abs/1904.03670

    2019 arXiv

  37. [45]

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

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Sonderegger, “Montreal Forced Aligner: Trainable Text- Speech Alignment Using Kaldi,” inInterspeech 2017. ISCA, Aug. 2017, pp. 498–502. [Online]. Available: https://www. isca-archive.org/interspeech 2017/mcauliffe17 intersp...

    2017

  38. [46]

    Morphosyntactic independence of STAMP morphs in Kono,

    A. Hamo, “Morphosyntactic independence of STAMP morphs in Kono,” 2025, conference presentation handout

    2025

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.