Pith. sign in

REVIEW 4 major objections 5 minor 45 references

From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Intermediate layers of a pre-trained wav2vec 2.0 model are the optimal representational space for learning articulatory mappings from speech alone, because they balance phonetic discriminability with speaker invariance.

desk verdict A useful layer-wise comparison of SSL representations for articulatory imitation, but the headline claims need statistical support and the synthesizer confound needs a control. read the letter →

arxiv 2509.05849 v2 pith:G2TIBYYJ submitted 2025-09-06 eess.AS

Marvin Lavechin , Thomas Hueber This is my paper
classification eess.AS
keywords acoustic-to-articulatorymappingself-supervisedlearningvocalimitationwav2vec2.0speakerinvariancearticulatorytrajectoriesplaceofarticulationspeechdevelopment
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

This paper asks how a learner with no access to another person's vocal tract can learn to imitate speech, and tests a possible answer in a self-supervised computational model. The model listens to a speech segment, maps its internal representation to articulatory parameters (jaw, tongue, lips), and synthesizes an imitation; only the mapping from representation to articulation is trained, by minimizing the distance between input and imitation in representational space. The central finding is that the representation matters: intermediate layers of a pre-trained wav2vec 2.0 model support human-like articulatory trajectories and intelligible imitated speech, while low-level MFCC features and very early or very late wav2vec layers do substantially worse, especially across speakers. The paper argues that what makes these intermediate layers effective is a balance between phonetic discriminability and speaker invariance, and reads this as computational evidence for the developmental proposal that perceptual phonetic learning precedes and guides articulatory development.

What carries the argument

The machinery is a perception-production loop with three frozen components and one trained component. A feature extractor (either MFCCs or the transformer layers of a pre-trained wav2vec 2.0 speech model) encodes input audio; a 2-layer bidirectional LSTM inverse model maps those representations to six articulatory parameters (jaw height, tongue body, tongue dorsum, tongue tip, lip protrusion, lip height); a pre-trained articulatory synthesizer, built from EMA recordings and guided PCA, turns those parameters plus source features into mel-spectrograms; and a HiFi-GAN vocoder renders speech. Only the inverse model is trained, by minimizing cosine distance between the input segment and the imit

What would settle it

Train the same model with the articulatory synthesizer replaced by a deliberately unrealistic or random vocal-tract mapping while keeping the wav2vec layer-8 representations; if articulatory trajectories still correlate with human EMA measurements and show high ABX place discrimination, then the synthesizer's fidelity is not necessary and the paper's explanation is incomplete. Conversely, evaluate the model on an independent EMA corpus from a different speaker: if the predicted trajectories do not correlate with that speaker's measured articulators, the single-speaker synthesizer is too narrow

Watch

Extended reading notes

Core claim

Central claim: a self-supervised imitation model can learn articulatory mappings from speech alone, and the representation used as the perceptual target controls success. Intermediate wav2vec 2.0 layers lift correlation with measured human articulatory trajectories from R≈0.37 (MFCCs) to R≈0.51 (single speaker) and place-of-articulation ABX from 74% to 80%. Across speakers, MFCC-trained models fail (56% ABX), while layer 8 reaches ≈77% ABX and word error rates near 15%. Probing shows phonetic information peaks at middle layers while speaker identity is minimized, and articulatory success tracks that balance. The paper concludes that phonetically informative, speaker-invariant representations

Load-bearing premise

The load-bearing premise is that the frozen articulatory synthesizer, learned from one French male speaker's EMA recordings, is a faithful and general enough model of the human vocal tract that optimizing through it teaches genuinely human-like articulation; if the synthesizer exploits artifacts, the learned 'articulatory trajectories' may not reflect real speech production.

Editorial extensions

If this is right

  • In a multi-speaker setting, low-level acoustic features (MFCCs) are insufficient for unsupervised articulatory learning; models trained on them produce trajectories that discriminate place of articulation worse than the raw acoustic signal does.
  • The optimal abstraction level shifts with task difficulty: layer 7 is best for imitating a single speaker, layer 8 for cross-speaker generalization, suggesting that more speaker-invariant representations become useful as input variability grows.
  • Because layer-wise performance is an inverted U, the same model can be used to map where phonetic versus speaker information lives in self-supervised speech representations; phonetic discriminability helps and speaker identifiability hurts articulatory learning.
  • The intelligibility result (WER ~15% at layer 7) shows that an inverse model trained with no articulatory labels and no explicit phonetic supervision can nevertheless reproduce linguistically identifiable speech through a frozen articulatory synthesizer.

Reading between the lines

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

  • A direct test of the balance hypothesis would be to take MFCC features and strip speaker identity from them (e.g., with an adversarial or normalization transformation) and measure whether multi-speaker ABX scores rise toward the wav2vec layer-8 level; the paper's correlation evidence predicts they would, but the paper does not run this experiment.
  • The same architecture could be run with layer-wise features from a model pretrained on French child-directed or noisy speech; if the optimal layer and its balance point shift, that would show the perceptual target is not universal but tuned to the learner's input ecology.
  • Because the synthesizer is fixed to one adult male vocal tract, the learned articulatory parameters cannot reflect infant tract growth. A version with a continuously morphing tract could test whether speaker-invariant representations also make inverse mappings robust to changing own-voice acoustics—a closer analogue to infant development.
  • The ABX place-of-articulation measure is computed in the articulatory parameter space; one could apply the same model to non-French speakers or languages to see if the 'optimal layer' shifts with phonological inventory, which would connect the mechanism to language-specific perceptual attunement.
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 / 5 minor

Summary. The paper presents a self-supervised computational model of acoustic-to-articulatory learning. A frozen wav2vec 2.0 feature extractor feeds a trained inverse model that predicts six articulatory parameters, which are then passed through a frozen articulatory synthesizer and vocoder to produce imitated speech. The inverse model is trained only by a cosine-similarity loss between input and output features. The authors compare MFCCs and wav2vec 2.0 layers in single-speaker (PB2009) and multi-speaker (Audiocite) settings, reporting that intermediate layers produce the best articulatory trajectories, place-of-articulation ABX scores, and intelligibility. Probing experiments are used to argue that intermediate layers optimally balance phonetic discriminability and speaker invariance. The paper interprets this as computational evidence for developmental theories in which perceptual learning of phonetic categories guides articulatory development.

Significance. If the central claim is correct, the paper makes a valuable contribution: it demonstrates an entirely unsupervised pathway from acoustic speech input to articulatory parameters, and it connects representation-learning properties to theories of speech development. The experimental architecture is cleanly designed: only the inverse model is trained, with the synthesizer and vocoder frozen, which enables targeted investigation of the mapping problem. The multi-metric evaluation (ground-truth correlation, ABX discrimination, intelligibility) and the inclusion of audio examples are strengths. However, the load-bearing conclusions currently rest on descriptive patterns and on a single-speaker articulatory synthesizer, so the significance is contingent on additional validation.

major comments (4)
  1. [§3.1.1, Fig. 2] The abstract and §3.1.1 claim that intermediate wav2vec 2.0 layers 'significantly outperform' MFCCs and yield a '38% improvement', but no significance tests are reported. The error bars in Fig. 2 represent standard deviations across 5 splits and appear to overlap substantially for the peak and neighboring layers (and possibly for MFCC at R≈0.37 vs. layer 6-7 at R≈0.51). The same issue affects Figs. 3, 4, and 6. Formal paired tests across splits or confidence intervals are needed to support the U-shaped claims and the superiority of intermediate layers over MFCC baselines.
  2. [§2.1, §2.4, §3.2] The multi-speaker conclusion may be an artifact of the single-speaker PB2009 synthesizer. In §3.2, Audiocite speakers (including females) are mapped onto articulatory parameters defined by a guided PCA of one French male speaker's EMA data entering the frozen synthesizer. The 'speaker invariance' of intermediate layers may therefore reflect specifically the ability to map diverse inputs onto this one vocal tract, not a general property of those representations. Since the synthesizer is fixed and is the only path from articulatory parameters to acoustics, changing it could shift the optimal layer or eliminate the advantage. A concrete test is to train or fine-tune the synthesizer on additional articulatory datasets or on a second vocal-tract model and check whether the inverted-U pattern and optimal layer are preserved.
  3. [§3.2.2, Fig. 5] The explanatory claim that 'phonetic discriminability plus speaker invariance drives articulatory learning' is based on layer-wise correlations between probe metrics and ABX scores computed across the same set of wav2vec 2.0 layers (n=13). No significance tests are reported, and the correlations may be confounded by layer depth and by the fact that the probe features and the features used to train the inverse models are identical representations. A partial correlation controlling for layer index, or a held-out validation in which probe-based predictions are made from untrained layers, would be needed to support the causal interpretation. As it stands, the correlation analysis is consistent with the claim but does not independently establish the proposed mechanism.
  4. [§3.2.3, Fig. 6] The intelligibility evaluation has a confound that weakens its support for the central claim. Source parameters (pitch period, pitch coefficient) are extracted directly from the input speech and passed unchanged through the synthesizer (§2.1). Since these source parameters already encode important phonetic and prosodic information, the WER pattern in Fig. 6 may partly reflect the source parameters rather than the learned articulatory trajectories. An ablation that predicts or degrades source parameters, or a comparison with a vocoder-only baseline, would clarify how much of the U-shaped intelligibility curve is due to the articulatory path.
minor comments (5)
  1. [§3.1.1] The text contains a placeholder link 'https://your-website.com'; this should be corrected to the actual audio examples page.
  2. [Fig. 3 caption] The caption describes the horizontal lines as 'ground truth articulatory parameters (blue) and MFCCs extracted directly from the input audio (yellow)', but the text in §3.1.2 calls the lower line 'orange'. The colors should be made consistent.
  3. [§3.2.2, Fig. 5] The right column of Fig. 5 plots correlations across layers, but the text does not report the correlation coefficients, p-values, or sample sizes. At minimum, these should be stated in the figure caption or main text.
  4. [§2.5] The ABX score description says 'd(A,X) < d(B,X)' and uses 'A and X are instances of the same consonant'; it would be clearer to specify whether X and B are also different contexts and whether the ABX task uses the 'same-different' or 'different-different' contrast convention. A brief clarification would improve reproducibility.
  5. [§2.4] The inverse model hyperparameters appear only in prose. A table summarizing feature dimensions, hidden sizes, learning rates, and training epochs for each component would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: layer comparison is empirical scanning with independent probes; self-citations are methodological and not load-bearing.

full rationale

The paper's central derivation is an empirical scan over frozen wav2vec 2.0 layers and MFCCs using a fixed PB2009-trained synthesizer. For each representation, a separate inverse LSTM is trained with the same cosine-distance objective, and performance is evaluated on held-out data: PB2009 test split for single-speaker correlation and ground-truth ABX, Audiocite test set for multi-speaker ABX and Whisper-transcribed WER, and LibriSpeech for phone/speaker probes. No equation is constructed so that the outcome equals an input; the 'balance' explanation is a post-hoc cross-layer correlation, not a fitted parameter renamed as a prediction. Self-citations (Georges et al. 2024 for guided PCA and ABX details; Badin et al. 2022 for the PB2009 corpus; Lavechin et al. 2024 as background) are methodological or motivational and are not load-bearing: the layer comparison and probes could be reproduced without them. The single-speaker synthesizer is a legitimate external-validity limitation, but it does not make the derivation circular. Therefore no circular step is identified.

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

The paper introduces no new physical or theoretical entities. The main assumptions are the correctness of the frozen synthesizer as a vocal tract model, the informativeness of wav2vec 2.0 features across layers, and the validity of the linear probes and cosine loss. The optimal layer index is effectively selected from test performance, which is the most notable free parameter.

free parameters (3)
  • Optimal wav2vec 2.0 layer index = 7 (single-speaker), 8 (multi-speaker)
    The 'optimal' layer is selected post-hoc as the layer with the highest ABX score on the test set (Figure 4). This layer choice is then used in Figure 5 and in the discussion as the representative balanced layer, making it a performance-fitted selection rather than a pre-registered prediction.
  • Inverse model learning rate = 1.7e-3
    Chosen by hand, not fitted to the central claim. Listed for completeness as a hyperparameter of the trained component.
  • Inverse model hidden size and layers = 64, 2-layer BiLSTM
    Chosen by hand, not varied or fitted. A standard configuration for this type of mapping.
assumptions (4)
  • domain assumption The articulatory synthesizer pre-trained on PB2009 is an accurate forward model of the vocal tract.
    Invoked throughout Sections 2.1 and 2.4. The inverse model learns entirely through this frozen synthesizer, so if the synthesizer's acoustic-to-articulatory mapping is biased or inaccurate, the learned articulatory parameters will be biased or inaccurate as well.
  • domain assumption wav2vec 2.0 representations, at any layer, contain enough information for the inverse mapping to be learned.
    The inverse model is trained to map these features to articulatory parameters through the synthesizer. If a layer lacked sufficient phonetic information, no training would succeed; the paper scans layers to find which contain usable information.
  • domain assumption Linear probes on wav2vec 2.0 layers accurately measure phonetic and speaker information content.
    Section 2.5 uses linear probes to claim that phone accuracy and speaker accuracy reflect the information available in each layer. This is a standard technique, but the correlation with articulatory learning outcomes (Figure 5) is descriptive and does not establish causation.
  • domain assumption Cosine distance in the chosen feature space is a valid objective for vocal imitation.
    The loss function minimizes cosine distance between input and imitated speech features. This assumes that closeness in the feature space corresponds to perceptual or functional equivalence, which is plausible for self-supervised features but not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model." pith.science (2026). https://pith.science/paper/G2TIBYYJ

@misc{pith2026250905849,
  author       = {Pith},
  title        = {Pith review of: From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2TIBYYJ}},
  note         = {Machine review of arXiv:2509.05849}
}
read the original abstract

Human infants face a formidable challenge in speech acquisition: mapping extremely variable acoustic inputs into appropriate articulatory movements without explicit instruction. We present a computational model that addresses the acoustic-to-articulatory mapping problem through self-supervised learning. Our model comprises a feature extractor that transforms speech into latent representations, an inverse model that maps these representations to articulatory parameters, and a synthesizer that generates speech outputs. Experiments conducted in both single- and multi-speaker settings reveal that intermediate layers of a pre-trained wav2vec 2.0 model provide optimal representations for articulatory learning, significantly outperforming MFCC features. These representations enable our model to learn articulatory trajectories that correlate with human patterns, discriminate between places of articulation, and produce intelligible speech. Critical to successful articulatory learning are representations that balance phonetic discriminability with speaker invariance -- precisely the characteristics of self-supervised representation learning models. Our findings provide computational evidence consistent with developmental theories proposing that perceptual learning of phonetic categories guides articulatory development, offering insights into how infants might acquire speech production capabilities despite the complex mapping problem they face.

Figures

Figures reproduced from arXiv: 2509.05849 by the authors.

Figure 1
Figure 1. Architecture of the speech imitation model. The system consists of four components: (1) a feature [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. shows correlations between predicted and ground-truth articulatory trajectories on PB2009. When using MFCC features to encode speech in￾put, the resulting articulatory predictions showed a moderate correlation with ground-truth parame￾ters (R ≃ 0.37). In contrast, wav2vec 2.0 embed￾dings as speech representations yielded substan￾tially stronger correlations in an inverted U-shaped pattern across layers. Peak perform… view at source ↗
Figure 3
Figure 3. Place of articulation ABX scores (%) obtained [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Left column: performance across wav2vec 2.0 layers (0-12) for place of articulation ABX (top), phone accuracy (middle), and speaker accuracy (bot￾tom). The red dashed line indicates layer 8, which yielded articulatory parameters most discriminative of place of articula…
Figure 6
Figure 6. Figure 6: Word Error Rate (%) of the imitated speech [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Examples of articulatory trajectories inferred by the model imitating in the MFCC or the wav2vec 2.0 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [1]

    J. David. Archibald, Charles A. Ferguson, Lise Menn, and Carol Stoel-Gammon. 1994. https://api.semanticscholar.org/CorpusID:109927262 Phonological development : models, research, implications . Language, 70:202

    work page 1994

  2. [2]

    Bishnu S Atal, Jih Jie Chang, Max V Mathews, and John W Tukey. 1978. Inversion of articulatory-to-acoustic transformation in the vocal tract by a computer-sorting technique. The Journal of the Acoustical Society of America, 63(5):1535--1555

    work page 1978

  3. [3]

    Pierre Badin, Gérard Bailly, Atef Ben Youssef, Frédéric Elisei, Christophe Savariaux, and Thomas Hueber. 2022. https://doi.org/10.5281/zenodo.6390598 PB2007 French acoustic-articulatory speech database . Zenodo, 6390598

    work page doi:10.5281/zenodo.6390598 2022

  4. [4]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449--12460

    2020

  5. [5]

    Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. 2023. Whisper X : Time-accurate speech transcription of long-form audio. Interspeech

    work page 2023

  6. [6]

    Marc Barnig. 2022. https://huggingface.co/datasets/mbarnig/lb-de-fr-en-pt-12800-TTS-CORPUS lb-de-fr-en-pt-12800-tts-corpus

    work page 2022

  7. [7]

    Ga s per Begu s , Alan Zhou, Peter Wu, and Gopala K Anumanchipalli. 2023. Articulation gan: Unsupervised modeling of articulatory learning. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE

    work page 2023

  8. [8]

    u rbis, Simon Stone, Patrick H \

    Peter Birkholz, Steffen K \"u rbis, Simon Stone, Patrick H \"a sner, R \'e mi Blandin, and Mario Fleischer. 2020. Printable 3d vocal tract shapes from mri data and their acoustic and aerodynamic properties. Scientific data, 7(1):255

    work page 2020

Show all 45 references
  1. [9]

    Catherine P Browman and Louis Goldstein. 1992. Articulatory phonology: An overview. Phonetica, 49(3-4):155--180

    1992

  2. [10]

    Cheol Jun Cho, Peter Wu, Abdelrahman Mohamed, and Gopala K Anumanchipalli. 2023. Evidence of vocal tract articulation in self-supervised learning of speech. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE

    2023

  3. [11]

    Cheol Jun Cho, Peter Wu, Tejas S Prabhune, Dhruv Agarwal, and Gopala K Anumanchipalli. 2024. Coding speech through vocal tract kinematics. IEEE Journal of Selected Topics in Signal Processing

    2024

  4. [12]

    Soline Felice, Sol \`e ne Evain, Solange Rossato, and Fran c ois Portet. 2024. Audiocite.net: A large spoken read dataset in F rench. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING)

    2024

  5. [13]

    Marc-Antoine Georges, Marvin Lavechin, Jean-Luc Schwartz, and Thomas Hueber. 2024. Decode, move and speak! S elf-supervised learning of speech units, gestures, and sound relationships using vocal imitation. Computational Linguistics, 50(4):1345--1373

    2024

  6. [14]

    Frank H Guenther. 1995. Speech sound acquisition, coarticulation, and rate effects in a neural network model of speech production. Psychological review, 102(3):594

    1995

  7. [15]

    Jean-R \'e my Hochmann and Liuba Papeo. 2014. The invariance problem in infancy: A pupillometry study. Psychological science, 25(11):2038--2046

    2014

  8. [16]

    Thomas Hueber, Eric Tatulli, Laurent Girin, and Jean-Luc Schwartz. 2020. https://doi.org/10.1162/neco_a_01264 Evaluating the potential gain of auditory and audiovisual speech-predictive coding using deep learning . Neural Computation, 32(3):596--625

    2020 doi

  9. [17]

    Susan S Jones. 2007. Imitation in infancy: The development of mimicry. Psychological science, 18(7):593--599

    2007

  10. [18]

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in neural information processing systems, 33:17022--17033

    2020

  11. [19]

    Patricia K Kuhl, Barbara T Conboy, Sharon Coffey-Corina, Denise Padden, Maritza Rivera-Gaxiola, and Tobey Nelson. 2008. Phonetic learning as a pathway to language: new data and native language magnet theory expanded ( NLM -e). Philosophical Transactions of the Royal Society B:...

    2008

  12. [20]

    perceptual magnet effect

    Patricia K Kuhl and Paul Iverson. 1995. Linguistic experience and the “perceptual magnet effect”. Speech perception and linguistic experience: Issues in cross-language research, pages 121--154

    1995

  13. [21]

    Patricia K Kuhl and Andrew N Meltzoff. 1996. Infant vocalizations in response to speech: Vocal imitation and developmental change. The journal of the Acoustical Society of America, 100(4):2425--2438

    1996

  14. [22]

    Patricia K Kuhl, Erica Stevens, Akiko Hayashi, Toshisada Deguchi, Shigeru Kiritani, and Paul Iverson. 2006. Infants show a facilitation effect for native language phonetic perception between 6 and 12 months. Developmental science, 9(2):F13--F21

    2006

  15. [23]

    Marvin Lavechin, Maureen De Seyssel, Lucas Gautheron, Emmanuel Dupoux, and Alejandrina Cristia. 2022. Reverse engineering language acquisition with child-centered long-form recordings. Annual Review of Linguistics, 8(1):389--407

    2022

  16. [24]

    Marvin Lavechin, Maureen de Seyssel, Marianne M \'e tais, Florian Metze, Abdelrahman Mohamed, Herv \'e Bredin, Emmanuel Dupoux, and Alejandrina Cristia. 2024. Modeling early phonetic acquisition from child-centered audio data. Cognition, 245:105734

    2024

  17. [25]

    Shinji Maeda. 1990. Compensatory articulation during speech: Evidence from the analysis and synthesis of vocal-tract shapes using an articulatory model. In Speech production and speech modelling, pages 131--149. Springer

    1990

  18. [26]

    Cl \'e ment Moulin-Frier, Sao M Nguyen, and Pierre-Yves Oudeyer. 2014. Self-organization of early vocal development in infants and machines: the role of intrinsic motivation. Frontiers in psychology, 4:1006

    2014

  19. [27]

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In International conference on acoustics, speech and signal processing (ICASSP), pages 5206--5210. IEEE

    2015

  20. [28]

    Martha Pelaez, Alexa R Borroto, and Jacqueline Carrow. 2018. Infant vocalizations and imitation as a result of adult contingent imitation. Behavioral Development, 23(1):81

    2018

  21. [29]

    Maxime Poli, Thomas Schatz, Emmanuel Dupoux, and Marvin Lavechin. 2024. Modeling the initial state of early phonetic learning in infants. Language Development Research, 5(1)

    2024

  22. [30]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning (ICML), pages 28492--28518. PMLR

    2023

  23. [31]

    Heikki Rasilo and Okko R \"a s \"a nen. 2017. An online model for vowel imitation learning. Speech Communication, 86:1--23

    2017

  24. [32]

    Denise E Ross and R Douglas Greer. 2003. Generalized imitation and the mand: Inducing first instances of speech in young children with autism. Research in Developmental Disabilities, 24(1):58--74

    2003

  25. [33]

    Clarence T Sasaki, Paul A Levine, Jeffrey T Laitman, and Edmund S Crelin. 1977. Postnatal descent of the epiglottis in man: a preliminary report. Archives of Otolaryngology, 103(3):169--171

    1977

  26. [34]

    Thomas Schatz, Vijayaditya Peddinti, Francis Bach, Aren Jansen, Hynek Hermansky, and Emmanuel Dupoux. 2013. Evaluating speech features with the minimal-pair ABX task: Analysis of the classical MFC / PLP pipeline. In Interspeech, pages 1--5

    2013

  27. [35]

    Antoine Serrurier, Pierre Badin, Anna Barney, Louis-Jean Bo \" e , and Christophe Savariaux. 2012. https://doi.org/10.1016/j.wocn.2012.08.001 The tongue in speech and feeding: Comparative articulatory modelling . Journal of Phonetics, 40(6):745--763

    2012 doi

  28. [36]

    Antoine Serrurier and Christiane Neuschaefer-Rube. 2023. Morphological and acoustic modeling of the vocal tract. The Journal of the Acoustical Society of America, 153(3):1867--1886

    2023

  29. [37]

    RE Stark. 1980. Stages of speech development in the first year of life. Child Phonology, 1

    1980

  30. [38]

    Changhan Wang, Morgane Rivi \`e re, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Miguel Pino, and Emmanuel Dupoux. 2021. https://api.semanticscholar.org/CorpusID:230433640 VoxPopuli : A large-scale multilingual speech corpus for representation lea...

    2021 arXiv

  31. [39]

    Anne S Warlaumont and Megan K Finnegan. 2016. Learning to produce syllabic speech sounds via reward-modulated neural plasticity. PloS one, 11(1):e0145096

    2016

  32. [40]

    Anne S Warlaumont, Gert Westermann, Eugene H Buder, and D Kimbrough Oller. 2013. Prespeech motor learning in a neural network using reinforcement. Neural Networks, 38:64--75

    2013

  33. [41]

    Janet F Werker and Richard C Tees. 1984. Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant behavior and development, 7(1):49--63

    1984

  34. [42]

    Yao-Yuan Yang, Moto Hira, Zhaoheng Ni, Artyom Astafurov, Caroline Chen, Christian Puhrsch, David Pollack, Dmitriy Genzel, Donny Greenberg, Edward Z Yang, and 1 others. 2022. Torchaudio: Building blocks for audio and speech processing. In International Conference on Acoustics, ...

    2022

  35. [43]

    Hyunjoo Yoo, Pumpki Lei Su, Gordon Ramsay, Helen L Long, Edina R Bene, and D Kimbrough Oller. 2024. Infant vocal category exploration as a foundation for speech development. Plos one, 19(5):e0299140

    2024

  36. [44]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  37. [45]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.