Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Enhancing Acoustic-to-Articulatory Speech Inversion by Incorporating Nasality

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single multi-task speech-inversion model that estimates oral tract variables, velum-related nasalance, and glottal source features together outperforms separate single-task models by about 5% and 9%.

desk verdict A competent multi-task speech-inversion paper whose headline 5%/9% gains are inflated by an embedding switch; the matched-embedding gains are the honest, modest core. read the letter →

arxiv 2506.09231 v1 pith:UZR6IGL6 submitted 2025-06-10 eess.AS

Saba Tabatabaee , Suzanne Boyce , Liran Oren , Mark Tiede , Carol Espy-Wilson This is my paper
classification eess.AS
keywords speechinversiontractvariablesnasalancevelopharyngealportmulti-tasklearningself-supervisedrepresentationsWavLM-Largesourcefeatures
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 proposes a unified speech-inversion system that recovers oral tract variables (lip and tongue constrictions), a velum-related nasalance measure, and three glottal source features from speech audio in one model. The central claim is that estimating these tracks together works better than estimating them separately: the multi-task system improves oral tract-variable estimation by about 5% relative and nasalance estimation by about 9% relative, compared with single-task baselines. The paper argues that nasality and oral constriction carry complementary information about the same articulatory gestures, so a shared representation lets each task help the other. If the claim holds, it means acoustic-to-articulatory inversion can be made more complete and more accurate by treating the whole vocal tract as one coordinated system, which matters for noninvasive study and clinical monitoring of velum function.

What carries the argument

The central object is a multi-task recurrent network over self-supervised speech representations. WavLM-Large embeddings are stacked, weighted by a 2D convolutional layer, and processed by three bidirectional GRU layers; two separate output heads then predict (a) the six oral tract variables and (b) the VP TV plus aperiodicity, periodicity, and F0. Tract variables (TVs) are synergies of articulator movement expressed as constriction degree and location coordinates; the VP TV is the nasalance measure, defined as $\mathrm{AE}_{\mathrm{nasal}}/(\mathrm{AE}_{\mathrm{nasal}}+\mathrm{AE}_{\mathrm{oral}})$, the ratio of nasal to nasal-plus-oral acoustic energy. The training loss is $\alpha(1-\mathrm{PC}) + (1-\alpha)\mathrm{RMSE}$ with $\alpha=0.8$, a weighted blend of Pearson correlation and root-mean-square error. This shared-encoder, two-head design is the proposed mechanism by which oral and nasal constriction information reinforce each other.

What would settle it

Train a WavLM-Large baseline on the same Nasometry-EGG splits as the MTL-SI model while keeping the oral-TV pipeline unchanged; if its VP TV Pearson correlation reaches roughly 0.9267 on unsegmented audio, the claimed synergy gain is an embedding effect rather than a multi-task effect. Separately, on a speaker with direct nasopharyngoscopy or real-time MRI of the velum, compare the MTL-SI VP TV trajectory against measured velum constriction; if the correlation is no higher than the teacher model's, the self-generated training target explains the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Multi-Task Learning Speech Inversion (MTL-SI) model can estimate six oral tract variables, a velopharyngeal-port tract variable (operationalized as nasalance), and three source features simultaneously, and that doing so improves all of them. The paper reports average oral-TV Pearson correlation of 0.8524 versus 0.8141 for the baseline system, a 4.70% relative gain, and VP TV Pearson correlations of 0.9267 (unsegmented audio) and 0.8852 (two-second segments) on the Nasometry-EGG test set, corresponding to relative gains of 5.82% and 9.08% over the earlier BiGRNN-HuBERT system. An ablation shows that removing either the source features or the VP TV from the multi-task setup lowers oral-TV performance, and a cross-corpus example ('Say packed memos', with simultaneous EMA and nasometry) shows the model reproducing the expected velum peaks for the two /m/ sounds. The paper concludes that consolidating VP TV and oral TV estimation into one model is more powerful than estimating them separately.

Load-bearing premise

The load-bearing premise is that the nasalance-based predictions produced by the authors' own earlier nasalance-inversion model are accurate enough to serve as training targets for velum movement; if that teacher model is biased, the combined system inherits the bias, and the reported gains could reflect self-consistency with the teacher rather than genuine recovery of velum articulation.

Editorial extensions

If this is right

  • A single speech-inversion model can output six oral tract variables, a nasalance-based velum track, and three source features at once, with both oral and nasalance estimates improving over single-task systems.
  • WavLM-Large self-supervised embeddings improve velum-related nasalance estimation compared with HuBERT-Large, extending earlier SSL findings to the velar domain.
  • The ablation indicates that removing either the source features or the VP TV lowers oral tract-variable accuracy, so nasalance and source information act as complementary inputs for oral estimation.
  • On an unseen corpus, the model reproduces the expected velum-opening peaks for two /m/ sounds, suggesting the joint model generalizes beyond its training corpora.
  • The ability to recover velum movement from audio alone points toward noninvasive monitoring of velopharyngeal port function, which the paper names as future clinical work.

Reading between the lines

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

  • Beyond the paper: the 5% oral-TV gain should be re-tested with direct velar kinematics, because the training target for the oral-TV experiments was itself generated by the authors' earlier nasalance-inversion model rather than by measured velum movement.
  • Beyond the paper: the reported 9% nasalance gain bundles a change in self-supervised embedding (WavLM-Large instead of HuBERT-Large) together with the multi-task structure, so an embedding-matched baseline is needed to isolate how much of the gain is synergy.
  • Beyond the paper: a clinical version of this system would need validation on disordered or atypical speech, since the current results come from healthy speakers reading standard passages.
  • Beyond the paper: the same shared-encoder logic could be tested on other articulatory pairings, such as tongue-root or pharyngeal constriction together with oral TVs, to see whether joint inversion helps whenever two constrictors are acoustically coupled.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a unified acoustic-to-articulatory speech inversion system that estimates oral tract variables (lip and tongue TVs) together with nasalance (VP TV) and glottal source features (periodicity, aperiodicity, F0) in a single model. Two training regimes are compared: single-task learning (STL-SI) and multi-task learning (MTL-SI), both using WavLM-Large self-supervised embeddings. The authors also develop an improved Nasal-SI system for VP TV estimation using WavLM-Large and compare it with a HuBERT-Large baseline. The paper reports that the MTL-SI system improves oral TV estimation by about 5% and nasalance estimation by about 9% relative to baseline models, and includes an ablation study and a qualitative cross-corpus example.

Significance. If the reported improvements are robust, the paper makes a useful contribution by demonstrating that oral and nasal/VP information can be estimated jointly in a speech inversion framework, with potential applications in clinical assessment. The paper provides a full experimental pipeline, a clear architecture comparison, and an ablation that helps identify the contribution of VP TV and source features. The use of WavLM-Large is a reasonable extension of prior work. However, the central quantitative claims are weakened by a confounded baseline comparison: the headline gains compare WavLM-Large MTL-SI against HuBERT-Large baselines, while the matched-embedding comparison inside the paper shows much smaller improvements. The small test sets and absence of significance testing further limit the strength of the conclusions.

major comments (3)
  1. [4.2.1, Table 3 and 4.2.2, Table 5] The headline results compare MTL-SI (WavLM-Large) against [16] and [6], which are HuBERT-Large systems. Table 2 shows that switching only the embedding from HuBERT-Large to WavLM-Large improves VP PPMC by 0.013-0.026 (e.g., 0.8533 to 0.8663 for segmented Nasal-SI), so a large part of the reported 9.08% relative gain (0.8115 to 0.8852 in the segmented condition) is attributable to the embedding rather than to multi-task consolidation. The matched-embedding comparison (MTL-SI vs STL-SI, both WavLM-Large) yields only about 0.6% relative VP gain (0.8852 vs 0.8802 segmented) and about 1.8% oral TV gain (0.8524 vs 0.8372 average). The authors should provide WavLM-Large versions of the [16] and [6] baselines, or at minimum explicitly decompose the gains into embedding and multi-task contributions, to support the abstract's 5%/9% claims.
  2. [3.2] VP TV ground truth for XRMB training is generated by the authors' own Nasal-SI system, so the STL-SI and MTL-SI models are trained to mimic that model's output for VP TV on XRMB. The evaluation on the Nasometry-EGG test set uses genuine nasalance labels and is therefore not circular, but the XRMB training signal cannot contain information beyond what the Nasal-SI system provides. The claim that the MTL framework 'enhances VP TV estimation' is consequently limited: the 5.82%/9.08% gains over [6] in Table 5 may reflect the better WavLM-Large embedding and architecture rather than the multi-task consolidation on XRMB. The paper should discuss this limitation and, if possible, validate the pseudo-labels on a held-out set.
  3. [Tables 2, 3, 5] No statistical significance is reported for any comparison, and the test sets are very small (2 speakers in Nasometry-EGG, 5 speakers in XRMB). The matched-embedding differences are small (PPMC differences around 0.005-0.015), so without confidence intervals, per-speaker results, or a significance test it is not possible to determine whether the multi-task gains are reliable. Please report per-speaker scores and a paired test (e.g., bootstrap over speakers) for the key comparisons.
minor comments (5)
  1. [Table 3] The column heading 'A VG. oral TVs' contains a typo and should read 'AVG. oral TVs'.
  2. [2.1] The test set consists of only 2 speakers; please specify whether the PPMC values in Tables 2 and 5 are averaged over the two speakers or pooled over utterances, as this affects interpretability.
  3. [Equation (2)] The loss function combines Pearson correlation and RMSE without explicit normalization; since the two terms have different scales, the alpha=0.8 weighting may be sensitive to target scaling. Please clarify how the terms are balanced, or report sensitivity to alpha.
  4. [4.2.3, Figure 4] The cross-corpus evaluation is qualitative and based on a single utterance. The generalization claim would be stronger with quantitative scores (e.g., PPMC between estimated and measured nasalance/EMA) on a larger set of utterances.
  5. [Abstract] The abstract's '5% in oral TVs estimation and 9% in nasalance estimation' should specify that these are relative improvements and identify the baselines (HuBERT-Large systems) to avoid implying matched-embedding gains.

Circularity Check

1 steps flagged · score 4.0 of 10

VP TV training targets for XRMB are generated by the authors' own Nasal-SI model, so the VP gains partly reflect distillation rather than independent recovery; oral TV results remain self-contained.

  1. fitted input called prediction [Section 3.2 (Developing STL-SI and MTL-SI Systems)]
    "We used the best-performing Nasal-SI system (with WavLM-Large embedding) to obtain ground truth estimates for VP TV for XRMB audios to develop the STL-SI and MTL-SI systems."

    The VP TV targets for XRMB are not direct measurements; they are outputs of the authors' own Nasal-SI model, which was itself trained on nasalance data. STL-SI and MTL-SI are trained to minimize a loss against these pseudo-labels, so their VP predictions on XRMB are, by construction, reproductions of Nasal-SI's outputs. Any systematic bias in Nasal-SI is inherited by the new models. The paper then claims improved VP TV estimation relative to separate baselines, but the VP training signal originates from the same model family, so the improvement is partly a distillation/embedding effect rather than independent recovery of velum movement.

full rationale

The strongest circularity concern is the pseudo-labeling of VP TV in Section 3.2: the ground truth for VP on XRMB is produced by the authors' own Nasal-SI model. This is a fitted input renamed as ground truth, and it makes the VP training loop partially circular. However, the paper's reported VP evaluations on the Nasometry-EGG test set use real nasalance measurements, so the headline VP numbers are not statistically forced by the training targets. The oral TV results are self-contained because they use direct X-ray pellet measurements converted via the same geometric transformation as the baseline [16]. The self-citations to [6] and [16] are not load-bearing in a circular way: [6] provides an external nasopharyngoscopy validation of nasalance, and [16] supplies a shared preprocessing transform. The additional concern that the 5% and 9% headline gains are confounded by switching from HuBERT-Large to WavLM-Large is a comparison-validity issue, not a circularity of derivation; the paper's own Table 2 shows the embedding upgrade accounts for much of the VP gain. Overall, the central oral-TV claim is independent, and the VP claim has independent test-set evaluation despite the circular training-label provenance, so a moderate score of 4 is appropriate.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several domain assumptions inherited from prior work, plus one ad hoc assumption of using model-generated pseudo-labels as ground truth for VP TV on XRMB. No new physical entities are introduced.

free parameters (2)
  • alpha in the loss function = 0.8
    In Equation 2, alpha balances Pearson correlation and RMSE; set empirically without sensitivity analysis.
  • Model hyperparameters (learning rate, batch size, hidden units, dropout) = LR 5e-4, batch size 8, 2x128 GRU for Nasal-SI, 2x512+256 GRU for STL/MTL
    Selected via grid search; these choices affect the reported PPMC scores.
assumptions (5)
  • domain assumption Nasalance is a valid ground truth proxy for velopharyngeal port constriction degree.
    The paper relies on the validation in [6] that compares nasalance with high-speed nasopharyngoscopy images of the velum.
  • domain assumption WavLM-Large and HuBERT-Large self-supervised embeddings contain sufficient articulatory information.
    Adopted from prior SSL literature and prior SI studies; not re-derived here.
  • domain assumption The geometric transformation from XRMB pellet trajectories to tract variables is valid.
    Borrowed from [16]; if the transformation is wrong, the oral TV ground truth would be wrong.
  • domain assumption The APP detector provides accurate estimates of periodicity, aperiodicity, and F0.
    These detector outputs are treated as ground truth for the source features without independent validation in this paper.
  • ad hoc to paper Pseudo-labels generated by the Nasal-SI model can serve as VP TV ground truth for XRMB audio.
    Section 3.2 explicitly uses the best Nasal-SI system to obtain VP TV estimates for XRMB; this is a circular setup because the generated labels are model outputs, not direct measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Acoustic-to-Articulatory Speech Inversion by Incorporating Nasality." pith.science (2026). https://pith.science/paper/UZR6IGL6

@misc{pith2026250609231,
  author       = {Pith},
  title        = {Pith review of: Enhancing Acoustic-to-Articulatory Speech Inversion by Incorporating Nasality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZR6IGL6}},
  note         = {Machine review of arXiv:2506.09231}
}
read the original abstract

Speech is produced through the coordination of vocal tract constricting organs: lips, tongue, velum, and glottis. Previous works developed Speech Inversion (SI) systems to recover acoustic-to-articulatory mappings for lip and tongue constrictions, called oral tract variables (TVs), which were later enhanced by including source information (periodic and aperiodic energies, and F0 frequency) as proxies for glottal control. Comparison of the nasometric measures with high-speed nasopharyngoscopy showed that nasalance can serve as ground truth, and that an SI system trained with it reliably recovers velum movement patterns for American English speakers. Here, two SI training approaches are compared: baseline models that estimate oral TVs and nasalance independently, and a synergistic model that combines oral TVs and source features with nasalance. The synergistic model shows relative improvements of 5% in oral TVs estimation and 9% in nasalance estimation compared to the baseline models.

Figures

Figures reproduced from arXiv: 2506.09231 by the authors.

Figure 1
Figure 1. A visual representation of vocal tract variables (adapted from [4]). production [5]. Nasality is controlled by constriction of the Velopharyngeal Port (VP), which regulates airflow and acoustic coupling between the nasal and oral cavities through the co￾ordinated movement of the velum and pharyngeal walls. Di￾rect observations of VP movement is invasive and costly be￾cause it requires trained specialists to administ… view at source ↗
Figure 2
Figure 2. Proposed model architecture for the Nasal-SI system. The system was initially developed using HuBERT-Large [24] from the SpeechBrain toolkit [25]. Subsequently, we de￾veloped the Nasal-SI system using WavLM-Large [26] em￾bedding. Both embeddings were evaluated to assess their im￾pact on the performance of the Nasal-SI system, with the best￾performing embedding model being proposed as the optimal in￾put. The Nasal-SI… view at source ↗
Figure 3
Figure 3. Proposed model architecture for the SI systems: A) STL-SI system B) MTL-SI system. STL-SI and MTL-SI were trained using the Adam opti￾mizer with a learning rate of 5e-4 and a batch size of 8, both selected via grid search. Early stopping with a patience of 8 epochs was used to prevent overfitting. The loss function to de￾velop STL-SI and MTL-SI is shown in Equation 2, which com￾bines Pearson Correlation (PC) and Roo… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Waveforms, spectrograms, followed by a comparison of the ground-truth EMA and nasalance with the correspond￾ing estimates from the MTL-SI system for the utterance “Say packed memos”. The second row shows the spectrogram, visu￾alizing the signal’s frequency content over…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages

  1. [16]

    Say packed memos

    HuBERT-Large - 0.8902 0.7142 0.7361 0.8180 0.8032 0.9229 - - - 0.8141 STL-SI WavLM-Large 0.9420 0.9026 0.7376 0.7591 0.8509 0.8307 0.9403 0.9326 0.8759 0.7551 0.8372 MTL-SI WavLM-Large 0.9462 0.9104 0.7594 0.7981 0.8626 0.8360 0.9478 0.9403 0.8815 0.7470 0.8524 of the proposed SI system design. In the subsequent steps, we estimated either the VP TV or 3SF...

    work page

  2. [6]

    Enhancing Acoustic-to-Articulatory Speech Inversion by Incorporating Nasality

    demonstrated that a BiGRNN-HuBERT system, leveraging HuBERT-Large SSL representations, outperforms a Temporal Convolutional Network (TCN) model using spectrogram inputs for VP TV estimation. The SSL-based system also exhibited su- perior generalization across various corpora, underscoring the robustness of SSL-derived features. Additionally, a compari- so...

    work page Pith review arXiv 2025

  3. [1]

    Introduction Speech articulation is a complex activity that requires finely timed coordination across articulators (lips, tongue, jaw, velum, and glottis) [1]. To recover this information from speech, a Speech Inversion (SI) system was developed that maps the speech signal not onto discrete articulator positions, but rather onto synergies of activity coor...

    work page

  4. [2]

    Dataset Description and Pre-Processing 2.1. Nasometry-EGG dataset In this study, we implemented a custom nasometry setup based on work [6], which computes nasalance as the relative acoustic energy from separate oral and nasal microphones mounted on a plate serving to isolate the two sources. Audio signals were recorded at a 51.2 kHz sampling rate. We coll...

    work page

  5. [3]

    Methodology 3.1. Developing Nasal-SI System We developed the Nasal-SI system to estimate VP TV from speech signals by integrating four parameters, including the EGG envelope (EGG-env) and three SFs, which are aperiodicity (Ap), periodicity (Per), and F0, based on research in [14]. The ground truth values for the three SF were extracted using the APP detec...

    work page

  6. [4]

    Results and Discussion 4.1. Results for Nasal-SI System To evaluate the Nasal-SI system, we used Pearson Product- Moment Correlation (PPMC) to measure the similarity between the ground truth and estimated parameters. Building on the ap- proach in [6], the training set was expanded by incorporating recordings from four additional speakers, to improve the V...

    work page

  7. [5]

    The results demonstrated that the proposed synergis- tic SI model improves the estimation of VP TV and oral TVs, underscoring the complementary nature of these characteris- tics

    Conclusions And Future Work In this paper, a novel SI system was proposed that simultane- ously estimates VP TV , along with oral TVs and three source features. The results demonstrated that the proposed synergis- tic SI model improves the estimation of VP TV and oral TVs, underscoring the complementary nature of these characteris- tics. Additionally, the...

    work page

  8. [7]

    H - 0.8757 0.8258 0.7309 0.5307 0.7644 Nasal-SI H - 0.8904 0.8331 0.7320 0.5348 0.7669 Nasal-SI W - 0.9152 0.8403 0.6683 0.5392 0.6957

    work page

Show all 37 references
  1. [8]

    Specifically, the relative improvements compared to the base- line BiGRNN-HuBERT are 6.75% and 4.51% for segmented and unsegmented audio conditions, respectively

    H 2 0.8115 0.8330 0.8373 0.8542 0.8562 Nasal-SI H 2 0.8533 0.8145 0.7648 0.5687 0.7214 Nasal-SI W 2 0.8663 0.8425 0.7673 0.6111 0.7209 As shown in Table 2, the Nasal-SI system with the WavLM- Large model outperforms alternative configurations, achieving the highest PPMC score ...

  2. [9]

    Streaming model for acoustic to articulatory inversion with transformer networks

    S. Udupa, A. Illa, and P. K. Ghosh, “Streaming model for acoustic to articulatory inversion with transformer networks.” inINTER- SPEECH, 2022, pp. 625–629

    2022

  3. [10]

    On the performance of ema-synchronized speech and stand-alone speech in acoustic-to-articulatory inversion,

    Q. Fang, “On the performance of ema-synchronized speech and stand-alone speech in acoustic-to-articulatory inversion,” inProc. Interspeech 2024, 2024, pp. 3110–3114

    2024

  4. [11]

    K. N. Stevens,Acoustic phonetics. MIT press, 2000, vol. 30

    2000

  5. [12]

    Towards an articulatory phonology,

    C. P. Browman and L. M. Goldstein, “Towards an articulatory phonology,”Phonology, vol. 3, pp. 219–252, 1986

    1986

  6. [13]

    A dynamical approach to gestural patterning in speech production,

    E. L. Saltzman and K. G. Munhall, “A dynamical approach to gestural patterning in speech production,”Ecological psychology, vol. 1, no. 4, pp. 333–382, 1989

    1989

  7. [14]

    Assessing neuromotor coordination in depression using inverted vocal tract variables

    C. Y . Espy-Wilson, A. C. Lammert, N. Seneviratne, and T. F. Quatieri, “Assessing neuromotor coordination in depression using inverted vocal tract variables.” inInterspeech, 2019, pp. 1448– 1452

    2019

  8. [15]

    Phonics: A large phoneme-grapheme frequency count revised,

    E. Fry, “Phonics: A large phoneme-grapheme frequency count revised,”Journal of Literacy Research, vol. 36, no. 1, pp. 85–98, 2004

    2004

  9. [17]

    Speaker-independent speech inversion for recovery of velopharyngeal port constriction de- gree,

    Y . M. Siriwardena, S. E. Boyce, M. K. Tiede, L. Oren, B. Fletcher, M. Stern, and C. Y . Espy-Wilson, “Speaker-independent speech inversion for recovery of velopharyngeal port constriction de- gree,”The Journal of the Acoustical Society of America, vol. 156, no. 2, pp. 1380–1390, 2024

    2024

  10. [18]

    Multi- corpus acoustic-to-articulatory speech inversion

    N. Seneviratne, G. Sivaraman, and C. Y . Espy-Wilson, “Multi- corpus acoustic-to-articulatory speech inversion.” inInterspeech, 2019, pp. 859–863

    2019

  11. [19]

    Speaker-independent acoustic-to-articulatory speech inversion,

    P. Wu, L.-W. Chen, C. J. Cho, S. Watanabe, L. Goldstein, A. W. Black, and G. K. Anumanchipalli, “Speaker-independent acoustic-to-articulatory speech inversion,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  12. [20]

    R. A. Krakow,The articulatory organization of syllables: A kine- matic analysis of labial and velar gestures. Yale University, 1989

    1989

  13. [21]

    Combining language corpora in a japanese electromagnetic articulography database for acoustic-to-articulatory inversion,

    T. Yan, K. Maekawa, Y . Nota, and M. Hirata, “Combining language corpora in a japanese electromagnetic articulography database for acoustic-to-articulatory inversion,” inProc. INTER- SPEECH, vol. 2023, 2023, pp. 1464–1467

    2023

  14. [22]

    Noise robust acoustic to articulatory speech inversion

    N. Seneviratne, G. Sivaraman, V . Mitra, and C. Y . Espy-Wilson, “Noise robust acoustic to articulatory speech inversion.” inInter- speech, 2018, pp. 3137–3141

    2018

  15. [23]

    Wav2nas: An exploratory approach to nasalance estimation in speech,

    R. Feng, Y .-A. Chen, Y .-L. Liu, J.-H. Yuan, and Z.-H. Ling, “Wav2nas: An exploratory approach to nasalance estimation in speech,” in2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2024, pp. 1–5

    2024

  16. [24]

    Subsequently, we de- veloped the Nasal-SI system using WavLM-Large [26] em- bedding

    from the SpeechBrain toolkit [25]. Subsequently, we de- veloped the Nasal-SI system using WavLM-Large [26] em- bedding. Both embeddings were evaluated to assess their im- pact on the performance of the Nasal-SI system, with the best- performing embedding model being proposed a...

  17. [25]

    Speaker-independent speech inversion for estimation of nasalance,

    Y . M. Siriwardena, C. Espy-Wilson, S. Boyce, M. K. Tiede, and L. Oren, “Speaker-independent speech inversion for estimation of nasalance,”arXiv preprint arXiv:2306.00203, 2023

    2023 arXiv

  18. [26]

    Ev- idence of vocal tract articulation in self-supervised learning of speech,

    C. J. Cho, P. Wu, A. Mohamed, and G. K. Anumanchipalli, “Ev- idence of vocal tract articulation in self-supervised learning of speech,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  19. [27]

    Improv- ing speech inversion through self-supervised embeddings and en- hanced tract variables,

    A. A. Attia, Y . M. Siriwardena, and C. Espy-Wilson, “Improv- ing speech inversion through self-supervised embeddings and en- hanced tract variables,” in2024 32nd European Signal Processing Conference (EUSIPCO). IEEE, 2024, pp. 306–310

    2024

  20. [28]

    The secret source: In- corporating source features to improve acoustic-to-articulatory speech inversion,

    Y . M. Siriwardena and C. Espy-Wilson, “The secret source: In- corporating source features to improve acoustic-to-articulatory speech inversion,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  21. [29]

    Motor speech disorders, saunders w,

    F. Darley, A. Aronsnon, and J. Brown, “Motor speech disorders, saunders w,”B, Philadelphia, pp. 171–975, 1975

    1975

  22. [30]

    Ieee recommended practice for speech quality measurements,

    E. H. Rothauser, “Ieee recommended practice for speech quality measurements,”IEEE Transactions on Audio and Electroacous- tics, vol. 17, no. 3, pp. 225–246, 1969

    1969

  23. [31]

    Speech production database user’s handbook,

    J. R. Westbury, “Speech production database user’s handbook,” IEEE Personal Communications-IEEE Pers. Commun., vol. 0, no, 1994

    1994

  24. [32]

    Masked autoencoders are articulatory learners,

    A. A. Attia and C. Y . Espy-Wilson, “Masked autoencoders are articulatory learners,” inICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  25. [33]

    Use of temporal information: Detection of periodicity, aperiodicity, and pitch in speech,

    O. Deshmukh, C. Y . Espy-Wilson, A. Salomon, and J. Singh, “Use of temporal information: Detection of periodicity, aperiodicity, and pitch in speech,”IEEE Transactions on Speech and Audio Processing, vol. 13, no. 5, pp. 776–786, 2005

    2005

  26. [34]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  27. [35]

    Speechbrain: A general-purpose speech toolkit,

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong et al., “Speechbrain: A general-purpose speech toolkit,”arXiv preprint arXiv:2106.04624, 2021

    2021 arXiv

  28. [36]

    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. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  29. [37]

    Acoustic- to-articulatory speech inversion with multi-task learning,

    Y . M. Siriwardena, G. Sivaraman, and C. Espy-Wilson, “Acoustic- to-articulatory speech inversion with multi-task learning,”arXiv preprint arXiv:2205.13755, 2022

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.