Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction

Arun Balaji Buduru; Girish; Mohd Mujtaba Akhtar; Orchid Chetia Phukan; Pailla Balakrishna Reddy; Rajesh Sharma; Swarup Ranjan Behera

REVIEW 4 major objections 6 minor 30 references

Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Speaker-recognition pre-trained models are the most effective features for singing-voice quality prediction, and fusing x-vector with ECAPA via the BATCH framework sets a new state of the art on the SingMOS benchmark.

desk verdict Useful empirical finding on speaker-recognition features for SingMOS, but the BATCH fusion loss is under-specified and the SOTA claim outruns the evidence. read the letter →

arxiv 2506.02232 v1 pith:R5IKNUD3 submitted 2025-06-02 eess.AS cs.SD

Orchid Chetia Phukan , Girish , Mohd Mujtaba Akhtar , Swarup Ranjan Behera , Pailla Balakrishna Reddy , Arun Balaji Buduru , Rajesh Sharma This is my paper · ORCID

classification eess.AScs.SD

keywords SingMOSpredictionsingingvoicequalityassessmentspeakerrecognitionpre-trainedmodelsx-vectorECAPABhattacharyyadistancefusionmodelself-supervisedspeechrepresentations

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that pre-trained speaker-recognition models—x-vector and ECAPA—extract the features most useful for predicting how human listeners rate synthesized singing voices, outperforming general speech self-supervised models and music pre-trained models. It tests this on the SingMOS benchmark with both fully connected and convolutional downstream heads, and the speaker-recognition models give the lowest errors on both test splits. The paper then introduces BATCH, a fusion method that aligns two pre-trained models' feature distributions with a Bhattacharyya-distance loss before concatenating them. Fusing x-vector and ECAPA with BATCH produces the best reported SingMOS scores, improving the CNN test-other1 MAE from 0.71 for the best single model to 0.14. If the result holds, singer-identity-oriented pretraining is a better source of fine-grained vocal cues for quality assessment than the usual speech- or music-representation targets.

What carries the argument

The load-bearing mechanism is BATCH (fusion via Bhattacharyya distance), which combines two frozen pre-trained representations through a gating operation $G(x)=\sigma(x)\odot x$, then aligns the two gated feature distributions with the Bhattacharyya distance $D_B(P,Q)=-\log\sum_x\sqrt{P(x)Q(x)}$ added as a loss term to mean squared error: $L = L_{\mathrm{MSE}} + \alpha L_{BD}$ with $\alpha=0.3$. The Bhattacharyya distance is the workhorse: unlike cosine similarity or KL divergence, it compares corresponding elements pointwise and is claimed to align heterogeneous feature spaces more stably, penalizing distant representations while encouraging compactness. After alignment the gated features are concatenated and fed to a 128-neuron fully connected regressor. BATCH adds only 2M-6M trainable parameters, so the gains come from the alignment loss, not from a large head.

What would settle it

Retrain the best BATCH model (x-vector + ECAPA, CNN head) on the SingMOS train split with shuffled MOS labels: if it still achieves test-other1 MAE near 0.14, the loss is not learning quality; alternatively, run the reported 0.14-MAE configuration over 10 random seeds and require the mean and worst-case MAE to stay far below the 0.71 single-model baseline before accepting the gain as generalization.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that models trained to recognize who is speaking generalize to judging how well someone sings. x-vector and ECAPA, both time-delay neural networks trained on VoxCeleb, consistently outperform monolingual, multilingual, and music PTMs with both FCN and CNN heads. The authors attribute this to speaker-recognition pretraining encoding pitch, tone, intensity, and rhythm—the same cues listeners use to judge synthesized singing quality. The second discovery is that the two speaker models are complementary: aligning their representations with the BATCH loss and concatenating them yields a CNN test-other1 MAE of 0.14, far below the best individual model's 0.71, and likewise tops test-main. The paper reports that this combination exceeds previous SingMOS methods.

Load-bearing premise

The claim rests on the assumption that BATCH's 2M-6M trainable parameters generalize from the few thousand SingMOS training clips rather than memorizing them; if the dramatic test-set drops reflect overfitting or accidental data-selection, the state-of-the-art claim collapses.

Editorial extensions

If this is right

Future SingMOS systems should include speaker-recognition embeddings such as x-vector and ECAPA as default baselines, since they beat both self-supervised speech models and music models in this study.
Fusing complementary pre-trained models with a distribution-alignment loss is a viable alternative to simple concatenation; BATCH beats concatenation across most of the 78 combinations tested.
The two speaker-recognition models encode complementary information, so their combination is not redundant; the paper reports this as the top-performing fusion.
Because the gains appear with both FCN and CNN heads on both test splits, the finding is not tied to one downstream architecture.

Reading between the lines

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

The magnitude of the reported gain (test-other1 CNN MAE 0.71 to 0.14) is unusually large for a small training set; one plausible explanation the paper does not test is that the Bhattacharyya loss acts as a strong regularizer on the 2M-6M trainable parameters, and ablating $\alpha$ would reveal how much of the gain is alignment versus regularization.
An extension the paper leaves implicit is applying BATCH to speech MOS or to non-intrusive assessment of other generative audio, where two complementary pretrained encoders could be aligned the same way.
The claim that speaker-identity pretraining is the key mechanism is inferred from model choice, not directly probed; one could test it by fine-tuning x-vector on a non-speaker task and checking whether the SingMOS advantage disappears.

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

Colleague,

You should know two things about this paper before deciding how much time to spend on it. The empirical comparison of individual pre-trained models is genuinely new and probably right: x-vector and ECAPA, both speaker-recognition embeddings, clearly beat the speech and music PTMs on both SingMOS test sets. That is a useful result for anyone building singing quality predictors. The second thing is that the paper's headline contribution—BATCH fusion with a Bhattacharyya-distance loss—is not mathematically well-defined as written, and the reported fusion gains are suspiciously large.

What the paper does well: it evaluates a wide range of PTMs (monolingual, multilingual, speaker-recognition, music) on the SingMOS benchmark, with both FCN and CNN downstream heads, and the ranking is consistent across test sets. The authors also correctly point out that the multilingual PTMs don't automatically win despite being exposed to Chinese and Japanese, which strengthens their claim that speaker-recognition pretraining captures something useful for singing voice.

The soft spots are real. First, the BD loss: L_BD = -log(sum sqrt(P(x)Q(x))) requires P and Q to be nonnegative and sum to 1. The paper never normalizes the gated features, and the gating G(x)=sigmoid(x)*x preserves the sign of x. If any product is negative, the square root is undefined; if all are positive but unnormalized, minimizing -log(sum) pushes magnitudes up, not toward alignment. As written, the objective cannot be the one that produced the numbers. This is load-bearing because the SOTA claim is specifically about BATCH.

Second, there is no comparison to the published SOTA numbers from Tang et al. or the ISCSLP paper. The authors say "setting SOTA" but only compare against their own re-runs, which is not the same as beating the best published result. Third, no error bars or seed variance on a dataset of only ~3,400 clips; the MAE drops from 0.71 to 0.14 on test-other1 are large enough to make me wonder about overfitting or evaluation artifacts. Fourth, the ablation in Table 2 removes both the gate and the BD term together, so you cannot tell which component causes the gain.

These issues are fixable. A revision with normalized features, a direct SOTA comparison, error bars, and a proper component ablation could turn this into a solid contribution. As it stands, the individual PTM finding is worth citing, but the BATCH claims need verification.

My recommendation: send it to peer review, not desk reject. The first-result on speaker-recognition PTMs is valuable enough that a careful referee can help the authors fix the technical flaws. I would bring it to a reading group only after they address the BD definition.

Referee Report

4 major / 6 minor

Summary. The paper evaluates thirteen pre-trained models for SingMOS prediction, including speaker-recognition embeddings (x-vector, ECAPA), self-supervised speech models (WavLM, Wav2vec2, Unispeech-SAT, XLS-R, Whisper, MMS), and music models (music2vec, MERT variants), using FCN and CNN heads. It reports that x-vector and ECAPA give the best individual results and that a new fusion method, BATCH, combining gated features with a Bhattacharyya-distance loss, achieves the best overall scores, claimed to be state of the art. The experiments use the SingMOS benchmark and compare with concatenation fusion.

Significance. The paper addresses an under-explored task and provides a broad, systematic comparison using public checkpoints and a public benchmark. If the findings are robust, the conclusion that speaker-recognition embeddings are strong features for singing-voice MOS is practically useful and could reorient feature selection for singing quality assessment. The authors are appropriately explicit that this is a hypothesis rather than a proven causal mechanism. However, the central BATCH contribution is not reproducible as written, and the SOTA claim lacks direct comparison; the paper is not yet ready to support its strongest conclusions.

major comments (4)

[3.2] The Bhattacharyya loss is undefined for the features produced by the described architecture. The gated features G(x)=σ(x)⊙x preserve the sign of x, and no non-negativity or normalization step is specified; therefore P(x) and Q(x) can take negative values, making sqrt(P(x)Q(x)) non-real, and if they are nonnegative but unnormalized, Σ_x sqrt(P(x)Q(x)) is not bounded by 1, so minimizing -log of that sum does not correspond to distribution alignment. Please specify a valid construction (e.g., softmax/probability normalization with a non-negativity guarantee) and verify that the reported numbers are produced by that construction; otherwise the BATCH results cannot be reproduced from the manuscript.
[4.2] The claim of state-of-the-art performance is not supported by any direct comparison. The paragraph 'Comparison to Methods used in Previous Works' names Tang et al. [9] and [10], but Table 2 reports only this paper's systems. Add a table with the published SingMOS MAE/MSE values for the prior methods on the same test splits, and state the selection criterion (e.g., best MAE or MSE) for the SOTA claim.
[4.1/4.2] All results are single-run with no seed variance or confidence intervals. The SingMOS training split is small (3,421 clips total, Section 4.1), while BATCH has 2M–6M trainable parameters, so the dramatic test-other1 CNN improvement from MAE 0.71 (x-vector) to 0.14 (x-vector+ECAPA) could reflect overfitting, early-stopping choices, or lucky initialization rather than generalization. Report results over multiple seeds with mean±std and describe the early-stopping and α-selection protocol; otherwise the magnitude of the fusion gain is not interpretable.
[3.2/Table 2] The concatenation baseline removes both the gating mechanism and the BD loss, so the reported BATCH-vs-concatenation gains cannot be attributed to the Bhattacharyya loss, the gating, or their interaction. Add ablations that include gate-only and BD-only variants (and, if applicable, the normalization step) to isolate the contribution of the proposed loss.

minor comments (6)

[Table 1] Table 1 lists 'Hu (HuBERT)' in the abbreviations but contains no HuBERT column; either include the HuBERT results or remove the mention in Section 2.
[Section 2] The parameter counts '94.70M, 94.68M, and 95.04M' are given for WavLM, Unispeech-SAT, HuBERT, and Wav2vec2, but only three numbers are provided for four models.
[Table 2] Table 2 contains two rows labeled 'MT95+MTV0' with different scores; one label is likely a typo and should be corrected.
[Throughout] The paper uses 'Bhattacharya' in several places; the conventional spelling is 'Bhattacharyya'.
[Abstract/3.2] There are minor language issues, including 'in a much more better way' in the abstract and 'continous' in Section 3.2; these should be corrected.
[Section 4.2/Table 1] MAE/MSE values are sometimes written with a '%' sign (e.g., '0.91%'); these are absolute errors, not percentages, and should be presented without the percent sign.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the SingMOS results are empirical benchmark comparisons with external checkpoints and held-out test splits.

full rationale

The paper's central claims are empirical. It freezes a set of publicly available pre-trained models (x-vector, ECAPA, WavLM, XLS-R, MERT, etc.), trains small FCN/CNN regressors on the official SingMOS train split, validates on the development split, and reports MSE/MAE on held-out test splits. The claim that speaker-recognition SPTMs are the best individual features, and that their BATCH fusion sets a new SOTA, is not derived from the hypothesis itself; it is supported by measurements against an external benchmark with external model checkpoints. The Bhattacharyya-distance loss in Section 3.2 raises a technical validity concern: raw gated features are not explicitly normalized or stated to be nonnegative probability distributions, so the objective as written may be ill-posed. That is a correctness/reproducibility issue, not circularity, because the loss is not defined in terms of the target MOS and the reported gain is an empirical outcome rather than an algebraic consequence of the loss formula. The only self-citation, reference [13], motivates the exploration of PTM fusion but is not load-bearing: the comparative results and the SOTA conclusion would stand or fall independently of that citation. No equation in the paper reduces to its own input, and no fitted parameter is relabeled as a prediction. Accordingly, the manuscript is not circular.

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

The central claim rests on the reliability of the SingMOS labels, the choice of frozen average-pooled features, the suitability of Bhattacharyya distance as an alignment loss, and the assumption that the small downstream model generalizes. None of these are demonstrated with independent evidence.

free parameters (2)

BD loss weight alpha = 0.3
Set after initial experimentation (Section 4.1, Training Details); its value affects the balance of L_MSE and L_BD and is not justified by a sensitivity analysis.
Best fusion pair selection = x-vector + ECAPA
Selected as the top performer among many pairwise combinations in Table 2; no multiple-testing correction, so this selection is effectively a fitted choice on the test sets.

assumptions (5)

domain assumption The MOS labels in the SingMOS dataset are reliable and treated as ground truth.
The paper trains regression models on these subjective scores without modeling annotator variance (Section 4.1).
domain assumption Frozen PTM features from the last hidden layer with average pooling preserve the vocal attributes relevant to SingMOS.
All PTMs are frozen and average-pooled (Section 2); no fine-tuning or layer-wise analysis is provided to support this.
ad hoc to paper Bhattacharyya distance is an appropriate alignment objective for PTM feature distributions.
Introduced as the core of BATCH (Section 3.2) with qualitative claims about stability, but no comparison against cosine, KL, or Euclidean alignment losses.
ad hoc to paper The hypothesis that speaker recognition pre-training equips the model to capture pitch, tone, and intensity.
Stated in Section 1 and Section 4.2 without a mechanistic analysis or controlled experiment.
domain assumption The small downstream model does not overfit the SingMOS training split.
BATCH has 2M to 6M trainable parameters trained on a few thousand clips with only early stopping and dropout (Section 4.1); no regularization analysis is presented.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction." pith.science (2026). https://pith.science/paper/R5IKNUD3

@misc{pith2026250602232,
  author       = {Pith},
  title        = {Pith review of: Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5IKNUD3}},
  note         = {Machine review of arXiv:2506.02232}
}

read the original abstract

In this study, we focus on Singing Voice Mean Opinion Score (SingMOS) prediction. Previous research have shown the performance benefit with the use of state-of-the-art (SOTA) pre-trained models (PTMs). However, they haven't explored speaker recognition speech PTMs (SPTMs) such as x-vector, ECAPA and we hypothesize that it will be the most effective for SingMOS prediction. We believe that due to their speaker recognition pre-training, it equips them to capture fine-grained vocal features (e.g., pitch, tone, intensity) from synthesized singing voices in a much more better way than other PTMs. Our experiments with SOTA PTMs including SPTMs and music PTMs validates the hypothesis. Additionally, we introduce a novel fusion framework, BATCH that uses Bhattacharya Distance for fusion of PTMs. Through BATCH with the fusion of speaker recognition SPTMs, we report the topmost performance comparison to all the individual PTMs and baseline fusion techniques as well as setting SOTA.

Figures

Figures reproduced from arXiv: 2506.02232 by the authors.

**Figure 1.** Proposed Framework: BATCH • We introduce a novel framework, BATCH for the fusion of PTMs. With BATCH, through the fusion of x-vector and ECAPA (Both are speaker recognition SPTMs), we report SOTA performance in SingMOS prediction, surpassing previous benchmarks and revealing new insights for future research. The full codebase and associated models will be made publicly available at: https://github.com/Helix-IIIT-D… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages

[9]

Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction

made the first contribution towards SingMOS prediction by introducing the first SingMOS prediction dataset, followed by a broader exploration of speech PTMs (SPTMs) for this task [10]. * Contributed equally as a first authors. Here, they have explored various SOTA SPTMs such XLS-R, wav2vec2, and so on that includes both monolingual and mul- tilingual SPTM...

work page Pith review arXiv 2025
[10]

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

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 enhancement,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–6

work page 2023
[1]

Introduction Mean Opinion Score (MOS) is a foundational metric for evaluat- ing audio quality, widely employed across speech and singing domains. Automatic MOS prediction offers numerous benefits, including efficiency and scalability, as it eliminates the need for time-intensive human evaluations and enables rapid assessment of large-scale samples generat...

work page
[2]

Speech PTMs: We use x-vector 1 [15] and ECAPA 2 [16] as speaker recognition SPTMs

Pre-Trained Models In this section, we explain the PTMs under consideration in our study. Speech PTMs: We use x-vector 1 [15] and ECAPA 2 [16] as speaker recognition SPTMs. Both are time delay neural net- work and trained on combination of V oxceleb1 + V oxceleb2. ECAPA improves over x-vector whereas x-vector improves over i-vector for speaker recognition...

work page
[3]

Modeling In this section, we discuss the downstream modeling with the PTMs followed by the proposed framework, BATCH for combina- tion of PTMs. 3.1. Individual Representation Modeling We use Fully Connected Network (FCN) and CNN as backend with individual PTMs. CNN comprises 1D convolutional layers with 64 and 128 filters, respectively with kernel size of...

work page
[4]

Dataset We utilize the sole SingMOS dataset by Tang et al

Experiments 4.1. Dataset We utilize the sole SingMOS dataset by Tang et al. [ 9]. It comprises of 3,421 singing voice clips in Chinese and Japanese, sampled at 16 kHz, with a total duration of approximately 4.25 hours. Each clip has an average length of 4.47 seconds. The dataset includes a diverse range of audio samples generated by 21 singing voice synth...

work page
[5]

We validated it by presenting comprehensive com- parative analysis of various SOTA PTMs including monolingual, multilingual, speaker recognition SPTMs as well as MPTMs

Conclusion In this study, we showed that the speaker recognition SPTMs such as x-vector, ECAPA are the most effective for SingMOS prediction. We validated it by presenting comprehensive com- parative analysis of various SOTA PTMs including monolingual, multilingual, speaker recognition SPTMs as well as MPTMs. Further, we proposed, BATCH for effective fusi...

work page
[6]

A study on incorporating whisper for robust speech as- sessment,

R. E. Zezario, Y .-W. Chen, S.-W. Fu, Y . Tsao, H.-M. Wang, and C.-S. Fuh, “A study on incorporating whisper for robust speech as- sessment,” in2024 IEEE International Conference on Multimedia and Expo (ICME), 2024, pp. 1–6

work page 2024

Show all 30 references

[7]

Somos: The samsung open mos dataset for the evaluation of neural text-to- speech synthesis,

G. Maniati, A. Vioni, N. Ellinas, K. Nikitaras, K. Klapsas, J. S. Sung, G. Jho, A. Chalamandaris, and P. Tsiakoulis, “Somos: The samsung open mos dataset for the evaluation of neural text-to- speech synthesis,” inInterspeech 2022, 2022, pp. 2388–2392

2022
[8]

The voicemos challenge 2023: zero-shot subjective speech quality prediction for multiple domains,

E. Cooper, W.-C. Huang, Y . Tsao, H.-M. Wang, T. Toda, and J. Yamagishi, “The voicemos challenge 2023: zero-shot subjective speech quality prediction for multiple domains,” in2023 IEEE Au- tomatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–7

2023
[11]

Investigating content-aware neural text-to-speech mos prediction using prosodic and linguistic fea- tures,

A. Vioni, G. Maniati, N. Ellinas, J. S. Sung, I. Hwang, A. Chala- mandaris, and P. Tsiakoulis, “Investigating content-aware neural text-to-speech mos prediction using prosodic and linguistic fea- tures,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and...

2023
[12]

Ensemble of deep neural network models for mos prediction,

M. Kunešová, J. Matoušek, J. Leheˇcka, J. Švec, J. Michálek, D. Ti- helka, M. Bulín, Z. Hanzlí ˇcek, and M. ˇRezáˇcková, “Ensemble of deep neural network models for mos prediction,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (IC...

2023
[13]

A non-intrusive speech quality assessment model using whisper and multi-head attention

G. Lin, Y . Tsao, and F. Chen, “A non-intrusive speech quality assessment model using whisper and multi-head attention.”
[14]

Indicmos: Multilingual mos prediction for 7 indian languages,

S. Udupa, S. Maiti, and P. K. Ghosh, “Indicmos: Multilingual mos prediction for 7 indian languages,” inInterspeech 2024, 2024, pp. 2690–2694

2024
[15]

Singmos: An extensive open- source singing voice dataset for mos prediction,

Y . Tang, J. Shi, Y . Wu, and Q. Jin, “Singmos: An extensive open- source singing voice dataset for mos prediction,”arXiv preprint arXiv:2406.10911, 2024

2024 arXiv
[16]

An exploration on singing mos prediction,

——, “An exploration on singing mos prediction,” in2024 IEEE 14th International Symposium on Chinese Spoken Language Pro- cessing (ISCSLP). IEEE, 2024, pp. 651–655

2024
[17]

Mos-bench: Benchmarking generalization abilities of subjective speech quality assessment models,

W.-C. Huang, E. Cooper, and T. Toda, “Mos-bench: Benchmarking generalization abilities of subjective speech quality assessment models,”arXiv preprint arXiv:2411.03715, 2024

2024 arXiv
[18]

Fusion of self-supervised learned models for mos prediction,

Z. Yang, W. Zhou, C. Chu, S. Li, R. Dabre, R. Rubino, and Y . Zhao, “Fusion of self-supervised learned models for mos prediction,” in Interspeech 2022, 2022, pp. 5443–5447

2022
[19]

Heterogeneity over homogeneity: Investigating multilingual speech pre-trained models for detecting audio deepfake,

O. Chetia Phukan, G. Kashyap, A. B. Buduru, and R. Sharma, “Heterogeneity over homogeneity: Investigating multilingual speech pre-trained models for detecting audio deepfake,” in Findings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Be...

2024
[20]

Investi- gation of ensemble features of self-supervised pretrained models for automatic speech recognition,

A. Arunkumar, V . Nileshkumar Sukhadia, and S. Umesh, “Investi- gation of ensemble features of self-supervised pretrained models for automatic speech recognition,” inInterspeech 2022, 2022, pp. 5145–5149

2022
[21]

X-vectors: Robust dnn embeddings for speaker recognition,

D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X-vectors: Robust dnn embeddings for speaker recognition,”2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5329–5333, 2018

2018
[22]

Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inInterspeech 2020, 2020, pp. 3830–3834

2020
[23]

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
[24]

Unispeech-sat: Universal speech repre- sentation learning with speaker aware pre-training,

S. Chen, Y . Wu, C. Wang, Z. Chen, Z. Chen, S. Liu, J. Wu, Y . Qian, F. Wei, J. Li, and X. Yu, “Unispeech-sat: Universal speech repre- sentation learning with speaker aware pre-training,”ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processin...

2022
[25]

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

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

2020
[26]

Xls-r: Self-supervised cross-lingual speech representation learning at scale,

A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Con- neau, and M. Auli, “Xls-r: Self-supervised cross-lingual speech representation learning at scale,” inInterspeech 2022, 2022, pp. 2278–2282

2022
[27]

Robust speech recognition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

2023
[28]

Scaling speech technology to 1,000+ languages,

V . Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel-Zarandiet al., “Scaling speech technology to 1,000+ languages,”Journal of Machine Learning Research, vol. 25, no. 97, pp. 1–52, 2024

2024
[29]

Map- music2vec: A simple and effective baseline for self-supervised music audio representation learning,

Y . Li, R. Yuan, G. Zhang, Y . Ma, C. Lin, X. Chen, A. Ragni, H. Yin, Z. Hu, H. He, E. Benetos, N. Gyenge, R. Liu, and J. Fu, “Map- music2vec: A simple and effective baseline for self-supervised music audio representation learning,”ArXiv, vol. abs/2212.02508, 2022

2022 arXiv
[30]

Mert: Acoustic music understanding model with large-scale self-supervised training,

Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetoset al., “Mert: Acoustic music understanding model with large-scale self-supervised training,”arXiv preprint arXiv:2306.00107, 2023

2023 arXiv

Pith tools

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

[1] [9]

Investigating the Reasonable Effectiveness of Speaker Pre-Trained Models and their Synergistic Power for SingMOS Prediction

made the first contribution towards SingMOS prediction by introducing the first SingMOS prediction dataset, followed by a broader exploration of speech PTMs (SPTMs) for this task [10]. * Contributed equally as a first authors. Here, they have explored various SOTA SPTMs such XLS-R, wav2vec2, and so on that includes both monolingual and mul- tilingual SPTM...

work page Pith review arXiv 2025

[2] [10]

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

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 enhancement,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–6

work page 2023

[3] [1]

Introduction Mean Opinion Score (MOS) is a foundational metric for evaluat- ing audio quality, widely employed across speech and singing domains. Automatic MOS prediction offers numerous benefits, including efficiency and scalability, as it eliminates the need for time-intensive human evaluations and enables rapid assessment of large-scale samples generat...

work page

[4] [2]

Speech PTMs: We use x-vector 1 [15] and ECAPA 2 [16] as speaker recognition SPTMs

Pre-Trained Models In this section, we explain the PTMs under consideration in our study. Speech PTMs: We use x-vector 1 [15] and ECAPA 2 [16] as speaker recognition SPTMs. Both are time delay neural net- work and trained on combination of V oxceleb1 + V oxceleb2. ECAPA improves over x-vector whereas x-vector improves over i-vector for speaker recognition...

work page

[5] [3]

Modeling In this section, we discuss the downstream modeling with the PTMs followed by the proposed framework, BATCH for combina- tion of PTMs. 3.1. Individual Representation Modeling We use Fully Connected Network (FCN) and CNN as backend with individual PTMs. CNN comprises 1D convolutional layers with 64 and 128 filters, respectively with kernel size of...

work page

[6] [4]

Dataset We utilize the sole SingMOS dataset by Tang et al

Experiments 4.1. Dataset We utilize the sole SingMOS dataset by Tang et al. [ 9]. It comprises of 3,421 singing voice clips in Chinese and Japanese, sampled at 16 kHz, with a total duration of approximately 4.25 hours. Each clip has an average length of 4.47 seconds. The dataset includes a diverse range of audio samples generated by 21 singing voice synth...

work page

[7] [5]

We validated it by presenting comprehensive com- parative analysis of various SOTA PTMs including monolingual, multilingual, speaker recognition SPTMs as well as MPTMs

Conclusion In this study, we showed that the speaker recognition SPTMs such as x-vector, ECAPA are the most effective for SingMOS prediction. We validated it by presenting comprehensive com- parative analysis of various SOTA PTMs including monolingual, multilingual, speaker recognition SPTMs as well as MPTMs. Further, we proposed, BATCH for effective fusi...

work page

[8] [6]

A study on incorporating whisper for robust speech as- sessment,

R. E. Zezario, Y .-W. Chen, S.-W. Fu, Y . Tsao, H.-M. Wang, and C.-S. Fuh, “A study on incorporating whisper for robust speech as- sessment,” in2024 IEEE International Conference on Multimedia and Expo (ICME), 2024, pp. 1–6

work page 2024