Multivariate Probabilistic Assessment of Speech Quality

Chandan K. A. Reddy; Christian Sch\"uldt; Fredrik Cumlin; Saikat Chatterjee; Victor Ungureanu; Xinyu Liang

REVIEW 4 major objections 4 minor 24 references

Multivariate Probabilistic Assessment of Speech Quality

T0 review · 4 major / 4 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 The paper claims a model can predict a proper five-dimensional Gaussian posterior over speech-quality dimensions, with covariance and correlations among MOS, noisiness, coloration, discontinuity, and loudness, while staying on par with…

desk verdict Clean, correct extension of probabilistic MOS to a multivariate Gaussian over five NISQA dimensions, but the paper's unique diagnostic claim rests on an untested Gaussian assumption and an affine transform typo. read the letter →

arxiv 2506.04890 v1 pith:WIRNDHGA submitted 2025-06-05 eess.AS

Fredrik Cumlin , Xinyu Liang , Victor Ungureanu , Chandan K. A. Reddy , Christian Schüldt , Saikat Chatterjee This is my paper

classification eess.AS

keywords speechqualityassessmentmeanopinionscoremultivariateGaussianposterioruncertaintyestimationCholeskydecompositionaffinetransformationnon-intrusiveMOSpredictionNISQAcorpus

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 replacing the single mean opinion score (MOS) with a full five-dimensional probabilistic score for speech clips, covering MOS along with noisiness, coloration, discontinuity, and loudness. It claims that a deep model can predict a proper multivariate Gaussian posterior over these five dimensions—mean vector plus covariance matrix—using a Cholesky parameterization and a generalized affine transformation, trained by maximum likelihood. The experimental claim is that this joint posterior matches state-of-the-art point-estimate accuracy on the NISQA validation sets while adding uncertainty and cross-dimension correlation information for free. That matters because a low MOS alone does not say what to fix; a predicted covariance between noise and overall quality points to the most effective remediation.

What carries the argument

The load-bearing object is the Cholesky transform: the network's head emits five mean values and fifteen entries of a lower-triangular matrix $L$; applying Softplus to $L$'s diagonal and forming $\mathrm{Softplus}_{\mathrm{diag}}(L) \mathrm{Softplus}_{\mathrm{diag}}(L)^T$ guarantees a symmetric positive-definite covariance for every input, and the parameterization is surjective. The generalized affine transform $(A\mu + b, A\Lambda A^T)$ with $A = 2I$ and $b = (3,3,3,3,3)$ maps predictions onto the $[-1,1]$ rating scale used during training, removing training bias. The Gaussian negative log-likelihood loss couples mean and covariance, so uncertainty and point estimates are learned jointly rather than as a post hoc calibration.

What would settle it

Compare the average of the predicted covariance matrices on a test set with the empirical covariance of residual score vectors computed by bootstrap-resampling the per-clip raters; if the 95% ellipsoids cover significantly fewer than 95% of the true score vectors, or the predicted correlations differ in sign from the bootstrap correlations, the Gaussian-posterior claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the five quality scores can be modeled as $\mathcal{N}(y; \hat{\mu}(x), \hat{\Lambda}(x))$, where $\hat{\mu}(x)$ is the mean vector and $\hat{\Lambda}(x)$ is the covariance matrix output by the network. The covariance is forced to be proper by writing $\hat{\Lambda} = \mathrm{Softplus}_{\mathrm{diag}}(L) \mathrm{Softplus}_{\mathrm{diag}}(L)^T$ with $L$ lower-triangular, and the output is affinely mapped by $(2\mu + 3, 4\Lambda)$ to the $[-1,1]$ rating scale. Trained with the Gaussian negative log-likelihood, the model's mean vector is on par with NISQA2, NISQA59, and Conformer on point metrics, while the covariance encodes the joint uncertainty and pairwise correlations among the five perceptual dimensions. The paper presents contour plots and correlation scatters showing that predicted correlations align with the intuitive pattern: when noisiness is poor relative to overall quality, noise and MOS are predicted strongly correlated, and when other distortions dominate, the correlation is low.

Load-bearing premise

The load-bearing premise is that, conditioned on the audio features, the five quality scores are jointly Gaussian; if the real conditional distribution is skewed, bounded, or multimodal, the predicted variances and correlations are biased.

Editorial extensions

If this is right

Per-clip confidence ellipses and correlation matrices become available for every quality prediction, so MOS estimates can be reported with calibrated uncertainty intervals.
When two dimensions are predicted strongly correlated, improving one is likely to lift the other; when they are not, the low score should be traced to a separate degradation.
The training recipe extends to any number of jointly rated quality dimensions, such as the three scores in ITU-T P.835, by changing only the output dimension.
Because point accuracy matches state-of-the-art point-estimate models, practitioners can adopt the probabilistic output without sacrificing ranking or regression performance.

Reading between the lines

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

A natural next step the paper does not take is to condition the predicted Gaussian on a low MOS to rank which quality dimension is most likely responsible; the covariance makes that conditional distribution one line of algebra away.
The fixed affine map ($A=2I$, $b=3$) is the only dataset-specific piece; the same head should transfer to other multi-rater corpora by re-estimating $A$ and $b$, which would be a direct test of the method's generality.
The predicted correlations could be checked empirically by bootstrap-resampling the 24-39 ratings per clip on the out-of-distribution test sets; the paper notes the difficulty but does not perform this validation.
If the conditional Gaussian assumption fails, the proper-covariance construction survives in a mixture or copula extension, but the reported correlations should then be read as linear associations rather than full conditional dependencies.

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.

Desk Editor's Note

read the letter

Worth a read if you work on non-intrusive speech quality. The paper does one thing cleanly: it replaces the scalar Gaussian posterior of DeePMOS/DNSMOS Pro with a five-dimensional Gaussian over NISQA's MOS/NOI/COL/DIS/LOUD dimensions, parameterizing the covariance via Cholesky with softplus diagonal and a generalized affine transform. The GNLL training is standard and the math is correct. Point estimates are on par with NISQA2 and Conformer, and they check three ablations (full covariance, independent variances, MSE head) over 10 runs. The code is public. That is a solid, reproducible contribution, and the subfield can use it.

The soft spot is exactly where the reader's stress test lands. The claimed added value is not the point estimates, it is the predicted covariance and the diagnostic story in Section 3.3.2. That claim is never validated. Training uses one mean-rating vector per clip, so the GNLL loss does not constrain the covariance to match any observable rater-level uncertainty. The authors acknowledge this in Section 3.3.2 and fall back on contour plots and a correlation-colored scatter plot. Those are suggestive, but they do not show that the predicted correlations are calibrated, or even that a Gaussian conditional is a better fit than a skewed or bounded alternative. The NISQA test sets have 24-39 ratings per clip, which is enough for at least a bootstrap covariance comparison on a subset, so the 'limited raters' excuse is weaker than it looks. The Gaussian assumption in Section 2.1 is load-bearing and untested.

Minor points: Eq. 5 says A = 2I with 'I is the unit 5x5 matrix,' but then b = (3,3,3,3,3). If the labels are 1-5, that maps to [-1,1], but the text says 'unit matrix' where it must mean identity, and the transformation as written should be A = 2I, b = -3 or the range is off. A typo, but it should be fixed. Also, Table 3 has odd formatting for NISQA TEST LIVETALK values, but the numbers seem internally consistent. Self-citation is present but not abusive; the multivariate extension is evaluated on external NISQA data and the baselines are fair.

Bottom line: the core method is sound, the point estimation is genuinely on par, and the paper is the first among the cited works to output a full multivariate posterior. The diagnostic claim needs either calibration evidence or a tempering of the language. This is a conference-quality paper that a serious referee should see, and I would engage with it rather than desk-reject. My verdict is conditional in the same sense as the reader's: accept if the covariance validation is added or the claims are scaled back.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes MultiGauss, a non-intrusive speech quality model that predicts a five-dimensional multivariate Gaussian posterior over MOS, noisiness, coloration, discontinuity, and loudness. The architecture freezes a wav2vec 2.0 feature extractor, trains a small head whose 20-dimensional output is split into a mean vector and a Cholesky factor, applies an affine transform, and is trained with a Gaussian negative log-likelihood loss (Eq. 8). The authors report point estimates that are on par with NISQA2 and Conformer across five NISQA datasets, and they argue that the predicted covariance matrix provides uncertainty and correlation estimates that enable better diagnosis of poor speech quality. The multivariate predictions are illustrated with two contour plots and a scatter plot colored by predicted correlation.

Significance. The central derivation is sound: the Cholesky parameterization guarantees a valid covariance matrix, and the GNLL objective is a standard maximum-likelihood formulation. The paper reports averages over 10 runs, releases source code, and evaluates across in-distribution validation sets and three out-of-distribution test sets. If the predicted covariance matrix were calibrated to rater-level dispersion, the method would offer a genuinely useful diagnostic tool for speech quality at no point-estimate cost. The current manuscript, however, does not demonstrate that calibration, and the diagnostic claims in Section 3.3.2 rest on an untested Gaussian assumption.

major comments (4)

[Section 3.3.2] The unique contribution of the paper is the predicted covariance matrix Lambda(x), but the manuscript never validates it against an observable quantity. The GNLL loss in Eq. (8) is optimized using one score vector per clip; a Lambda that reduces training loss can be arbitrarily uncalibrated as a conditional covariance. The text explicitly states that 'due to limited raters, validating multivariate predictions with methods like bootstrapping is difficult' and substitutes two qualitative contour plots and a scatter plot. This is insufficient for the central diagnostic claim. Please add a quantitative check, for example: on the test sets with 24-39 ratings per clip, compute per-clip empirical covariance matrices (or bootstrap covariances) of the rater scores and compare them with the predicted Lambda(x), e.g., by reporting the correlation or mean squared error between predicted and empirical covariance entries, or by evaluating coverage of predicted 95% ellipsoids. A comparison against the MultiGauss indep baseline would show whether the correlations carry information.
[Section 2.1] The Gaussian posterior is adopted 'following DNSMOS Pro [4]' and 'motivated by the analytical tractability,' but this assumption is load-bearing for the inferences in Section 3.3.2. Quality scores are bounded on [1,5] and are likely skewed or heteroscedastic; under misspecification, Lambda(x) does not equal the conditional covariance. Please test the assumption against a non-Gaussian alternative or provide calibration evidence: for instance, report the empirical coverage of marginal predictive intervals at the clip level, or compare GNLL on a held-out set against a model with independent variances or with beta margins. If such a test is not feasible within the current scope, the diagnostic interpretation of Lambda(x) should be explicitly softened.
[Section 2.2.2, Eq. (5)] The affine transformation is stated to be (A mu + b, A Lambda A^T) with A=2I and b=(3,3,3,3,3), and the text says this 'transforms the quality labels onto the scale [-1,1].' Applied to ratings in [1,5], this map produces [5,13], so the direction of the transformation is inconsistent with the text. Either the labels are first normalized to [-1,1] and the affine map returns predictions to the original scale, or the affine parameters are misreported. Because Eq. (8) uses y_n without specifying whether the labels are normalized, this ambiguity prevents reproduction of the training procedure. Please state the preprocessing of y_n and the exact role of the affine transform.
[Table 3] The NISQA TEST LIVETALK columns for NISQA2 and NISQA59 appear to have RMSE and PCC transposed. As printed, NISQA2 has RMSE 0.829 and PCC 0.684, while MultiGauss has RMSE 0.502 and PCC 0.834; this contradicts the paper's statement that MultiGauss is 'on par with NISQA2 and Conformer' and is inconsistent with the relative magnitudes in Tables 1 and 2. Please verify the table and re-state the comparison after correction.

minor comments (4)

[Section 2.2.1] The text contains the typo 'Cholesky transfrom' instead of 'Cholesky transform'; in addition, Eq. (4) would be clearer if written as Lambda = \tilde{L} \tilde{L}^T, where \tilde{L} is obtained from L by applying the Softplus operation to the diagonal.
[Section 2.2] The selection of wav2vec 2.0 layer 12 is described only as 'based on a preliminary study'; please report the selection procedure or explicitly note that this is a tuned hyperparameter, since it is the only free architectural choice reported.
[Figure 3] The color gradient encodes predicted correlation, but no color bar is provided; please add one so the reader can interpret the magnitudes and sign of the correlation values.
[Section 3.1] The manuscript should state how the five per-dimension ratings are aggregated into the vector y_n used in Eq. (8), since rater-level data are otherwise not used in training or evaluation.

Circularity Check

1 steps flagged · score 2.0 of 10

Transparent self-citation supplies the Gaussian ansatz, but no equation-level circularity; covariance outputs are genuine predictions.

ansatz smuggled in via citation [Section 2.1 (Problem formulation), Eqs. (1)-(6)]
"Following DNSMOS Pro [4], we model the posterior as a Gaussian motivated by the analytical tractability. Since there are five quality labels, we will model the posterior with a multivariate Gaussian distribution."

The paper's distinctive output—joint uncertainty and correlation estimates—depends on the multivariate Gaussian form of p(y|x). The only cited justification for this form is [4], the authors' own DNSMOS Pro, which introduced the same Gaussian ansatz for a univariate posterior. This paper neither derives conditional Gaussianity nor tests it against rater-level covariance or non-Gaussian baselines; Section 3.3.2 explicitly declines validation of multivariate predictions. The central probabilistic premise is therefore imported from a self-citation rather than independently established. The issue is limited because the Gaussian choice is openly declared as a tractability-motivated modeling assumption, and the point-estimation architecture is evaluated on external NISQA data.

full rationale

The paper's derivation chain is largely self-contained beyond the Gaussian assumption. The Cholesky parameterization (Eq. 4) and affine transform (Eq. 5) are standard mathematical constructions; the GNLL loss (Eq. 8) is a textbook maximum-likelihood objective. The predicted mean vector and covariance matrix are outputs of the trained network, not fitted constants renamed as predictions, and the point-estimate performance is compared against external NISQA test sets. The only load-bearing step that leans on prior work is the Gaussian ansatz, explicitly adopted from the authors' own DNSMOS Pro paper; because that prior work itself adopts the Gaussian form without deriving it, the ansatz is imported via self-citation. The paper does not validate the covariance predictions against rater-level data, and Section 3.3.2 acknowledges this. This makes the diagnostic claims dependent on an untested modeling choice, but the central probabilistic framework is not circular in an equation-by-equation sense: the model's outputs (μ(x), Λ(x)) are genuinely predicted and not predetermined by the training objective. One minor self-citation that is not load-bearing for the point-estimation results, and a partially load-bearing self-citation for the Gaussian form, lead to a score of 2 rather than 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The model introduces no new physical or conceptual entities. It rests on a modeling choice (Gaussianity), a dataset assumption (NISQA label reliability), and a fixed affine transform whose description contains a typo. The only hand-chosen hyperparameter is the wav2vec2 layer.

free parameters (1)

wav2vec2 feature layer index = 12
Chosen via a preliminary study (Section 2.2) rather than derived or cross-validated on all folds; affects all five output dimensions.

assumptions (3)

domain assumption The five quality scores are jointly Gaussian given audio features.
Invoked in Section 2.1 to define p(y|x) = N(mu(x), Lambda(x)); this tractability assumption underlies the GNLL loss and the interpretation of the covariance as uncertainty.
domain assumption The NISQA corpus provides reliable ground-truth ratings for all five dimensions.
All training, validation, and tests use NISQA labels; if the labels are noisy or biased, the estimated covariance structure inherits that noise.
ad hoc to paper The affine transform with A=2I, b=(3,3,3,3,3) maps correctly between model output space and the 1-5 rating scale.
Eq. (5) is presented as a fixed transform; however the accompanying text states it maps labels to [-1,1], which contradicts the algebra. The reported results imply the transform is applied as written, but the text is inconsistent.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multivariate Probabilistic Assessment of Speech Quality." pith.science (2026). https://pith.science/paper/WIRNDHGA

@misc{pith2026250604890,
  author       = {Pith},
  title        = {Pith review of: Multivariate Probabilistic Assessment of Speech Quality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIRNDHGA}},
  note         = {Machine review of arXiv:2506.04890}
}

read the original abstract

The mean opinion score (MOS) is a standard metric for assessing speech quality, but its singular focus fails to identify specific distortions when low scores are observed. The NISQA dataset addresses this limitation by providing ratings across four additional dimensions: noisiness, coloration, discontinuity, and loudness, alongside MOS. In this paper, we extend the explored univariate MOS estimation to a multivariate framework by modeling these dimensions jointly using a multivariate Gaussian distribution. Our approach utilizes Cholesky decomposition to predict covariances without imposing restrictive assumptions and extends probabilistic affine transformations to a multivariate context. Experimental results show that our model performs on par with state-of-the-art methods in point estimation, while uniquely providing uncertainty and correlation estimates across speech quality dimensions. This enables better diagnosis of poor speech quality and informs targeted improvements.

Figures

Figures reproduced from arXiv: 2506.04890 by the authors.

**Figure 1.** Overview of the architecture. where ψ are the parameters of the posterior distribution. Following DNSMOS Pro [4], we model the posterior as a Gaussian motivated by the analytical tractability. Since there are five quality labels, we will model the posterior with a multivariate Gaussian distribution. This means the problem is to estimate ψ(x) = (µ(x), Λ(x)), since pψ (y|x) = N (y; µ(x), Λ(x)). Consider a DNN as a r… view at source ↗

**Figure 2.** Contour plots of the marginal distributions over MOS and NOI [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Scatter plot of the MOS and NOI labels on the NISQA VAL SIM dataset. The correlation values are the predicted correlation values by the MultiGauss model. Furthermore, the MultiGauss model demonstrates consistent performance, with low standard deviation across evaluation metrics and quality dimensions, indicating the robustness of its architecture and optimization objective. As a generalizability test, we also make… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages

[4]

Our method gen- eralizes existing maximum likelihood frameworks and affine transformations to a multivariate context

Conclusion We have proposed a novel approach to probabilistic non- intrusive speech quality assessment by predicting a multivariate Gaussian distribution of speech quality scores. Our method gen- eralizes existing maximum likelihood frameworks and affine transformations to a multivariate context. We have made archi- tectural considerations to predict a pr...

work page
[1]

Subjective SQA is done by let- ting several raters rate the quality of the speech clip based on an ordinal scale such as the mean-opinion-score (MOS) scale [1]

Introduction Speech quality assessment (SQA) is the task of assessing the speech quality of speech clips. Subjective SQA is done by let- ting several raters rate the quality of the speech clip based on an ordinal scale such as the mean-opinion-score (MOS) scale [1]. Subjective SQA is considered the paramount method for eval- uating speech quality, but is ...

work page
[2]

Multivariate Probabilistic Assessment of Speech Quality

Method 2.1. Problem formulation Letxdenote the features of a speech clip. For each speech clip, five quality scores are given to quantify five different quality aspects, denoted byy= (y 1, ..., y5). A multivari- ate (non-intrusive) speech quality dataset is given byD= {(xn yn)}N n=1, whereNis the total number of speech clips. Instead of modeling the point...

work page Pith review arXiv 2025
[3]

Experiments 3.1. Datasets We use the NISQA Corpus [17] for both training and evaluation, which includes two training datasets (NISQA TRAIN SIM, NISQA TRAIN LIVE), two validation datasets (NISQA V ALSIM, NISQA V ALLIVE), and three test datasets (NISQA TEST LIVETALK, NISQA TEST P501, NISQA TEST FOR). Each speech clip is rated on five quality aspects: overal...

work page
[5]

W. Lin, D. Tao, J. Kacprzyk, Z. Li, E. Izquierdo, and H. Wang, Multimedia Analysis, Processing and Communications. New York: Springer Publishing, 2011

work page 2011
[6]

Deepmos: Deep posterior mean-opinion-score of speech,

X. Liang, F. Cumlin, C. Sch ¨uldt, and S. Chatterjee, “Deepmos: Deep posterior mean-opinion-score of speech,” inInterspeech

work page
[7]

Deepmos-b: Deep posterior mean-opinion-score using beta distribution,

X. Liang, F. Cumlin, V . Ungureanu, C. K. Reddy, C. Sch¨uldt, and S. Chatterjee, “Deepmos-b: Deep posterior mean-opinion-score using beta distribution,” in2024 32nd European Signal Process- ing Conference (EUSIPCO). IEEE, 2024, pp. 416–420

work page 2024
[8]

Dnsmos pro: A reduced-size dnn for probabilistic mos of speech,

F. Cumlin, X. Liang, V . Ungureanu, C. KA Reddy, C. Sch¨uldt, and S. Chatterjee, “Dnsmos pro: A reduced-size dnn for probabilistic mos of speech,” inProc. Interspeech 2024, 2024, pp. 4818–4822

work page 2024

Show all 24 references

[9]

Subjective test methodology for evaluating speech communication systems that include noise suppression algo- rithm

I.-T. R. P.835, “Subjective test methodology for evaluating speech communication systems that include noise suppression algo- rithm.” 2003

2003
[10]

MOSNet: Deep learning-based ob- jective assessment for voice conversion,

C.-C. Lo, S.-W. Fu, W.-C. Huang, X. Wang, J. Yamagishi, Y . Tsao, and H.-m. Wang, “MOSNet: Deep learning-based ob- jective assessment for voice conversion,” inInterspeech 2019, 09 2019, pp. 1541–1545

2019
[11]

LDNet: Unified listener dependent modeling in MOS prediction for syn- thetic speech,

W.-C. Huang, E. Cooper, J. Yamagishi, and T. Toda, “LDNet: Unified listener dependent modeling in MOS prediction for syn- thetic speech,” inICASSP 2022 - 2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2022

2022
[12]

Latent-based neural net for non-intrusive speech quality assessment,

F. Cumlin, C. Sch ¨uldt, and S. Chatterjee, “Latent-based neural net for non-intrusive speech quality assessment,” in2023 33th Eu- ropean Signal Processing Conference (EUSIPCO), ser. European Signal Processing Conference, sep 2023, pp. 36–40

2023
[13]

DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise sup- pressors,

C. Reddy, V . Gopal, and R. Cutler, “DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise sup- pressors,” inIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6-11, 2021, 06 2021

2021
[14]

Utmos: Utokyo-sarulab system for voicemos challenge 2022,

T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “Utmos: Utokyo-sarulab system for voicemos challenge 2022,” inProc. Interspeech 2022, 09 2022, pp. 4521– 4525

2022
[15]

The ZevoMOS entry to V oiceMOS Challenge 2022,

A. Stan, “The ZevoMOS entry to V oiceMOS Challenge 2022,” inProc. Interspeech 2022, 2022, pp. 4516–4520. [Online]. Available: https://www.isca-speech.org/archive/pdfs/ interspeech 2022/stan22 interspeech.pdf

2022
[16]

Le-ssl- mos: Self-supervised learning mos prediction with listener en- hancement,

Z. Qi, X. Hu, W. Zhou, S. Li, H. Wu, J. Lu, and X. Xu, “Le-ssl- mos: Self-supervised learning mos prediction with listener en- hancement,” in2023 IEEE International Conference on Multime- dia and Expo (ICME). IEEE, 2023

2023
[17]

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
[18]

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

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

2020
[19]

Generaliza- tion ability of mos prediction networks,

E. Cooper, W.-C. Huang, T. Toda, and J. Yamagishi, “Generaliza- tion ability of mos prediction networks,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8442–8446

2022
[20]

Dnsmos p. 835: A non-intrusive perceptual objective speech quality metric to eval- uate noise suppressors,

C. K. Reddy, V . Gopal, and R. Cutler, “Dnsmos p. 835: A non-intrusive perceptual objective speech quality metric to eval- uate noise suppressors,” inICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 886–890

2022
[21]

NISQA: A deep CNN-self-attention model for multidimensional speech quality prediction with crowdsourced datasets,

G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “NISQA: A deep CNN-self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” inInterspeech
[22]

Probing speech quality information in asr systems,

B. T. Ta, M. T. Le, N. M. Le, and V . H. Do, “Probing speech quality information in asr systems,” inProceedings of INTER- SPEECH, 2023, pp. 541–545

2023
[23]

Reformulation strategies for eigenvalue optimization using sylvester’s criterion and cholesky decomposition,

D. S. Wicaksono and W. Marquardt, “Reformulation strategies for eigenvalue optimization using sylvester’s criterion and cholesky decomposition,” inComputer Aided Chemical Engineering. El- sevier, 2013, vol. 32, pp. 487–492

2013
[24]

Adam: A method for stochastic opti- mization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,”arXiv preprint arXiv:1412.6980, 2014

2014 arXiv

Pith tools

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

[1] [4]

Our method gen- eralizes existing maximum likelihood frameworks and affine transformations to a multivariate context

Conclusion We have proposed a novel approach to probabilistic non- intrusive speech quality assessment by predicting a multivariate Gaussian distribution of speech quality scores. Our method gen- eralizes existing maximum likelihood frameworks and affine transformations to a multivariate context. We have made archi- tectural considerations to predict a pr...

work page

[2] [1]

Subjective SQA is done by let- ting several raters rate the quality of the speech clip based on an ordinal scale such as the mean-opinion-score (MOS) scale [1]

Introduction Speech quality assessment (SQA) is the task of assessing the speech quality of speech clips. Subjective SQA is done by let- ting several raters rate the quality of the speech clip based on an ordinal scale such as the mean-opinion-score (MOS) scale [1]. Subjective SQA is considered the paramount method for eval- uating speech quality, but is ...

work page

[3] [2]

Multivariate Probabilistic Assessment of Speech Quality

Method 2.1. Problem formulation Letxdenote the features of a speech clip. For each speech clip, five quality scores are given to quantify five different quality aspects, denoted byy= (y 1, ..., y5). A multivari- ate (non-intrusive) speech quality dataset is given byD= {(xn yn)}N n=1, whereNis the total number of speech clips. Instead of modeling the point...

work page Pith review arXiv 2025

[4] [3]

Experiments 3.1. Datasets We use the NISQA Corpus [17] for both training and evaluation, which includes two training datasets (NISQA TRAIN SIM, NISQA TRAIN LIVE), two validation datasets (NISQA V ALSIM, NISQA V ALLIVE), and three test datasets (NISQA TEST LIVETALK, NISQA TEST P501, NISQA TEST FOR). Each speech clip is rated on five quality aspects: overal...

work page

[5] [5]

W. Lin, D. Tao, J. Kacprzyk, Z. Li, E. Izquierdo, and H. Wang, Multimedia Analysis, Processing and Communications. New York: Springer Publishing, 2011

work page 2011

[6] [6]

Deepmos: Deep posterior mean-opinion-score of speech,

X. Liang, F. Cumlin, C. Sch ¨uldt, and S. Chatterjee, “Deepmos: Deep posterior mean-opinion-score of speech,” inInterspeech

work page

[7] [7]

Deepmos-b: Deep posterior mean-opinion-score using beta distribution,

X. Liang, F. Cumlin, V . Ungureanu, C. K. Reddy, C. Sch¨uldt, and S. Chatterjee, “Deepmos-b: Deep posterior mean-opinion-score using beta distribution,” in2024 32nd European Signal Process- ing Conference (EUSIPCO). IEEE, 2024, pp. 416–420

work page 2024

[8] [8]

Dnsmos pro: A reduced-size dnn for probabilistic mos of speech,

F. Cumlin, X. Liang, V . Ungureanu, C. KA Reddy, C. Sch¨uldt, and S. Chatterjee, “Dnsmos pro: A reduced-size dnn for probabilistic mos of speech,” inProc. Interspeech 2024, 2024, pp. 4818–4822

work page 2024