Pith. sign in

REVIEW 3 major objections 6 minor 41 references

SAMOS: A Neural MOS Prediction Model Leveraging Semantic Representations and Acoustic Features

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read A speech quality predictor that fuses semantic and acoustic features beats prior state-of-the-art on the BVCC benchmark.

desk verdict Competent MOS prediction system paper with real but small gains; overclaims 'significant' without error bars. read the letter →

arxiv 2411.11232 v1 pith:ECI4U3LQ submitted 2024-11-18 cs.SD eess.AS

classification cs.SDeess.AS
keywords MOSpredictionspeechqualityassessmentsemanticrepresentationacousticfeaturewav2vec2BiVocodermulti-tasklearningBVCC
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 authors propose SAMOS, a neural model that predicts the naturalness of synthesized speech (mean opinion score, MOS) by combining two complementary views: semantic representations from a pretrained wav2vec2 model and acoustic features from a pretrained BiVocoder feature extractor. They claim this fusion, together with listener-ID embeddings and parallel regression/classification heads, achieves state-of-the-art system-level correlation on the BVCC main track of the VoiceMOS Challenge 2022 using a single network, and remains comparable to ensemble baselines on the out-of-domain BC2019 set. If correct, this means that incorporating explicit acoustic detail alongside SSL semantics improves MOS prediction without needing expensive ensembles.

What carries the argument

The central mechanism is the pairing of a semantic stream (wav2vec2 embeddings) with an acoustic stream (BiVocoder feature extractor followed by a Conformer), aligned frame-by-frame under the assumption that the two extractors use the same frame shift. The model then applies a stage-wise training schedule: first a regression head with clipped and contrastive losses, then a classification head with cross-entropy, then a separately trained linear aggregation layer that combines the two heads' scores. The claim is that each component is necessary: ablations removing the semantic module, acoustic module, ID embedding, weight branch, regression head, classification head, or aggregation layer all degrade at least one metric on BVCC.

What would settle it

Compute the frame counts and frame shift of wav2vec2 and BiVocoder features on the same utterance; if they differ, or if temporally shifting one stream by a single frame changes BVCC system-level SRCC by more than a small amount, the alignment assumption is violated. Alternatively, train SAMOS with acoustic features replaced by a random one-frame shift and compare system-level SRCC on BVCC.

Watch

Extended reading notes

Core claim

SAMOS jointly uses semantic and acoustic information: wav2vec2 produces frame-level semantic embeddings, BiVocoder's feature extractor produces a compressed low-dimensional feature containing both amplitude and phase, and a Conformer captures global context. These are concatenated with a learned listener embedding and fed to a BiLSTM-based predictor with two heads—a regression head using weighted frame aggregation and a classification head over score distributions—whose outputs are fused by a learned aggregation layer. The paper reports that on the BVCC test set SAMOS achieves the highest system-level LCC (0.944) and SRCC (0.942) among the compared models, including ensembles, and lowers MSE relative to SSL-MOS; on BC2019 it is second on ranking metrics only to the much larger UTMOS ensemble.

Load-bearing premise

The paper assumes that wav2vec2 and BiVocoder produce features with identical frame timing, so concatenating their frame-level features aligns matching moments of speech; if that alignment is imperfect, the fused representation may combine mismatched context.

Editorial extensions

If this is right

  • SAMOS provides a stronger single-model baseline for MOS prediction on English synthetic speech, outperforming prior single-model and some ensemble systems on BVCC system-level correlation.
  • On out-of-domain Mandarin data (BC2019), SAMOS remains competitive with the ensemble UTMOS and surpasses UTMOS strong on most metrics.
  • The ablation results indicate that semantic and acoustic streams each carry information the other lacks; removing either hurts performance.
  • The multi-task regression/classification framework with a learned aggregation layer improves accuracy over either head alone.
  • Because no ensemble is used, SAMOS is computationally lighter than top VoiceMOS entries while achieving comparable ranking quality.

Reading between the lines

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

  • The frame-alignment assumption could be tested by measuring whether SAMOS is sensitive to small temporal offsets between semantic and acoustic streams; if robust, the method could extend to other SSL feature pairs.
  • The acoustic stream's contribution suggests that phase information, which most MOS models discard, is perceptually relevant; this points toward phase-aware metrics for synthesis evaluation.
  • The listener-ID embedding and the virtual mean-listener trick might transfer to other subjective rating tasks where raters differ, such as video quality or aesthetic assessment.
  • The stage-wise training and aggregation layer could be applied to any two complementary feature types, not just semantic/acoustic, as long as frame alignment holds.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes SAMOS, a MOS prediction model that combines semantic representations from a frozen wav2vec2, acoustic features from a frozen BiVocoder feature extractor, and listener embeddings. These features are concatenated and fed into a BiLSTM-based predictor with parallel regression and classification heads, trained in three stages, followed by an aggregation layer. Evaluated on the BVCC main track and the BC2019 out-of-domain track, SAMOS reports the best system-level LCC, SRCC, and KTAU among the compared models on BVCC, and is described as 'comparable' on BC2019. The central claim is that fusing semantic and acoustic features with multi-task heads and an aggregation layer yields state-of-the-art single-model MOS prediction.

Significance. If the result holds, the paper demonstrates that combining SSL-derived semantic features with explicit acoustic features (amplitude and phase from a vocoder feature extractor) and listener information improves MOS prediction over SSL-only baselines. Strengths of the study include the use of external human MOS labels, frozen pretrained extractors, held-out test sets, and a systematic ablation study that isolates each component. The architectural idea is sensible and the empirical results are promising, but the evidence for a statistically reliable improvement over prior systems is incomplete, which limits the strength of the conclusions as currently stated.

major comments (3)
  1. [4.1 / Table 1] The statement in §4.1 that SAMOS "significantly outperformed baseline models" on BVCC is not supported by the reported evidence. The margins over UTMOS/T11 are 0.003–0.005 in S-LCC, S-SRCC, and S-KTAU, while SAMOS has higher S-MSE (0.097) than UTMOS (0.090) and DDOS (0.091). No significance tests, confidence intervals, or multiple-seed variance are provided, despite §3.3 acknowledging randomness in checkpoint selection. Since system-level correlations are computed over a finite set of test systems, a 0.003–0.005 gap may be within sampling noise. Please add statistical significance testing (e.g., bootstrap over systems or listeners, or repeated runs with different seeds) or temper the claims to "the best reported correlation among the compared models on BVCC for these metrics."
  2. [2.2] The fusion design concatenates semantic and acoustic features along the time axis, based on the assertion that "the frame shift settings of wav2vec and BiVocoder are the same, so the number of frames for semantic and acoustic features is identical." The paper does not report the actual frame shift/hop sizes for either extractor or empirically verify that the frame indices correspond to the same temporal locations. If the extractors' internal downsampling rates differ or there is a constant offset, the concatenation mixes misaligned context, which could affect the results. Please provide the frame alignment details or a verification experiment to support this load-bearing design assumption.
  3. [4.1 / Table 1, BC2019] The abstract's claim of "comparable performance" on BC2019 is overstated. On BC2019, SAMOS trails UTMOS on all four metrics and the S-MSE gap is large (0.179 vs. 0.030); it also has higher S-MSE than T11 (0.048) and DDOS (0.070). While the comparison with the single-model UTMOS strong is more favorable, the overall statement should be qualified. Please report single-model comparisons explicitly and state which metrics support the "comparable" characterization.
minor comments (6)
  1. [Throughout] There are several typos: "Serveral" in §2 and §2.2.1, "incorproating" in the Introduction, "abaltion" in the Table 2 caption, and "listenr" in the §2.1.1 heading.
  2. [3.3] The sentence "The output feature dimensions of the semantic and acoustic modules were 64 and 768, respectively" appears reversed relative to typical wav2vec2 output dimensions (768) and compressed acoustic features (often lower-dimensional). Please verify the dimensions.
  3. [2.2 / Figure 2] Figure 2, the three-stage training diagram, is difficult to follow; consider labeling the fixed and trainable modules more clearly or adding a textual description of what changes between stages 1 and 2.
  4. [3.3] The sentence "If the system-level SRCC didn't decrease within 15 epochs, early stopping was applied" is ambiguous: early stopping is usually triggered when the metric does not improve, not when it does not decrease. Please clarify the intended condition.
  5. [2.2.1] The contrastive loss is defined for a pair (xi, xj), but the paper does not specify how pairs are sampled or how the loss is aggregated over a batch. Please provide this detail for reproducibility.
  6. [3.1] The sentence "Since the raters in BVCC and BC2019 datasets are different, we defaulted to considering the listeners as mean ones, when finetuning on BC2019" could be clarified: does this mean only the mean-listener ID is used and individual rater scores are ignored during fine-tuning?

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SAMOS's features come from externally pretrained extractors, targets are human MOS labels, and evaluation is on held-out test sets.

full rationale

The paper's claimed derivation chain is not circular. The MOS targets come from external human ratings (BVCC/BC2019), the semantic features come from a frozen fairseq wav2vec2 pretrained on unlabeled speech, and the acoustic features come from the frozen feature extractor of BiVocoder, which is the authors' prior work but was trained for waveform reconstruction on VCTK, not for MOS prediction. The prediction network is trained on the BVCC training split and evaluated on the held-out test split; baselines are external challenge systems. The only self-citation is [24] BiVocoder; because it is a fixed, MOS-label-independent feature extractor with its own external training objective and because its contribution is measured by ablations on held-out data, this citation is not load-bearing circularity. The assertion that wav2vec2 and BiVocoder share the same frame shift is an empirical alignment premise, not a definitional equivalence; if incorrect it would be a correctness risk, not a circularity. Likewise, the absence of significance tests or confidence intervals for the 0.003-0.005 metric gaps is a statistical-evidence weakness, not a reduction of the prediction to its inputs. No equation in the paper defines a target in terms of the model's own output, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 8 free parameters · 4 assumptions · 1 invented entities

SAMOS is an empirical system built from pretrained extractors and hand-set hyperparameters. The central performance claim relies on transferability of pretrained features, exact frame alignment, and trust in human MOS labels. No new physical entities are introduced; the only invented concept is a virtual mean-listener ID, which is an inference-time approximation.

free parameters (8)
  • contrastive loss margin alpha = 0.1
    Hand-set hyperparameter in the contrastive loss Lcon; no sensitivity analysis is reported. Section 2.2.1.
  • regression loss weight beta = 1
    Hand-set weight for the clipped loss; no sensitivity analysis is reported. Section 2.2.1.
  • contrastive loss weight gamma = 0.5
    Hand-set weight for the contrastive loss; no sensitivity analysis is reported. Section 2.2.1.
  • clipped loss threshold tau = 0.25
    Hand-set threshold in the clipped loss to reduce overfitting; no sensitivity analysis is reported. Section 2.2.1.
  • BiLSTM hidden size = 128
    Architecture choice in the base MOS predictor; no search is reported. Section 3.3.
  • number of best checkpoints averaged = 3
    Parameter averaging choice affects the final predictions; no sensitivity analysis is reported. Section 3.3.
  • learning rate = 0.0001
    Optimization setting chosen without reported tuning. Section 3.3.
  • batch size = 8
    Optimization setting chosen without reported tuning. Section 3.3.
assumptions (4)
  • domain assumption Human MOS annotations in BVCC and BC2019 are reliable ground truth for naturalness.
    The model is trained and evaluated against these labels; if they are noisy or biased, all metrics inherit that noise. Section 3.1.
  • domain assumption Pretrained wav2vec2 and BiVocoder representations remain informative for utterances in BVCC and BC2019.
    Semantic and acoustic modules are used as frozen feature extractors; their transfer quality is not separately evaluated. Sections 2.1.1 and 2.1.2.
  • domain assumption Frame counts of wav2vec2 and BiVocoder features are identical so they can be concatenated directly.
    The paper states that frame shifts are the same, but gives no empirical verification of alignment. Section 2.2.
  • ad hoc to paper A virtual mean-listener ID can represent unknown raters at inference.
    The mean-listener ID is introduced to handle the absence of rater IDs at test time; its validity is only tested indirectly through system-level metrics. Section 2.1.1.
invented entities (1)
  • virtual mean-listener ID
    purpose: Represents the average listener during training and serves as the rater input at inference when the true rater is unknown.
    It is a modeling artifact with no external falsifiable handle; its contribution is inferred only from ablations and final metrics. Section 2.1.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAMOS: A Neural MOS Prediction Model Leveraging Semantic Representations and Acoustic Features." pith.science (2026). https://pith.science/paper/ECI4U3LQ

@misc{pith2026241111232,
  author       = {Pith},
  title        = {Pith review of: SAMOS: A Neural MOS Prediction Model Leveraging Semantic Representations and Acoustic Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECI4U3LQ}},
  note         = {Machine review of arXiv:2411.11232}
}
read the original abstract

Assessing the naturalness of speech using mean opinion score (MOS) prediction models has positive implications for the automatic evaluation of speech synthesis systems. Early MOS prediction models took the raw waveform or amplitude spectrum of speech as input, whereas more advanced methods employed self-supervised-learning (SSL) based models to extract semantic representations from speech for MOS prediction. These methods utilized limited aspects of speech information for MOS prediction, resulting in restricted prediction accuracy. Therefore, in this paper, we propose SAMOS, a MOS prediction model that leverages both Semantic and Acoustic information of speech to be assessed. Specifically, the proposed SAMOS leverages a pretrained wav2vec2 to extract semantic representations and uses the feature extractor of a pretrained BiVocoder to extract acoustic features. These two types of features are then fed into the prediction network, which includes multi-task heads and an aggregation layer, to obtain the final MOS score. Experimental results demonstrate that the proposed SAMOS outperforms current state-of-the-art MOS prediction models on the BVCC dataset and performs comparable performance on the BC2019 dataset, according to the results of system-level evaluation metrics.

Figures

Figures reproduced from arXiv: 2411.11232 by the authors.

Figure 1
Figure 1. Overall structure of the proposed SAMOS model, where the “concat” represents the feature concatenation operation, p represents the probability scores of various classes outputted by the classification head, r and c represent the scores outputted by the regression head and classification head respectively, and s represents the final score. 2. Proposed Method The model structure of SAMOS is illustrated in [PITH_FULL_… view at source ↗
Figure 2
Figure 2. Training process of the proposed SAMOS. 2.1.2. Acoustic module The acoustic module consists of a BiVocoder and a Conformer. BiVocoder, which is our previous work [24], is a newly pro￾posed bidirectional neural vocoder with both feature extrac￾tion and waveform generation capabilities Regarding the fea￾ture extraction module, speech amplitude and phase spectra are separately passed through a ConvNeXt v2 network [29] … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [1]

    Common ob- jective evaluation metrics, such as mel-cepstral distance (MCD)

    Introduction Text-to-speech (TTS) synthesis and voice conversion (VC) are focal points in the field of speech research, where the evaluation of the quality of speech synthesized by TTS and VC systems in- volves both objective and subjective assessments. Common ob- jective evaluation metrics, such as mel-cepstral distance (MCD)

  2. [2]

    As a result, some ob- jective measures or models related to human perception have been proposed [3, 4, 5, 6]

    and signal-to-noise ratio (SNR) [2], have limited correlation with human perception of speech quality. As a result, some ob- jective measures or models related to human perception have been proposed [3, 4, 5, 6]. However, these objective methods typically require reference speech, making it impractical for evaluating synthesized speech signals. Therefore,...

  3. [3]

    Dataset In this paper, the experiments followed the same settings as the V oiceMOS Challenge 2022 [15]

    Experiment Setup 3.1. Dataset In this paper, the experiments followed the same settings as the V oiceMOS Challenge 2022 [15]. The dataset includes BVCC dataset [12] from the main track and BC2019 dataset [27] from the out-of-domain (OOD) track. The BVCC dataset contains 7,106 English utterances, with the training/development/test sets split in a ratio of ...

  4. [4]

    mean-listener

    Proposed Method The model structure of SAMOS is illustrated in Figure 1. First, the feature extractor produces three types of features, i.e., se- mantic representations derived from a wav2vec2-based seman- tic module [23], acoustic features extracted by an acoustic mod- ule composed of the feature extractor of BiV ocoder [24] and Conformers [28], and list...

  5. [5]

    When the rater ID is not the mean-listener, the label representing the sample is the score given by the individual rater (an integer i from 1 to 5)

    Assuming the vector is [p1, p2, p3, p4, p5], the classification score is the expectation c = P5 i=1 i × pi. When the rater ID is not the mean-listener, the label representing the sample is the score given by the individual rater (an integer i from 1 to 5). In this case, the target is a one-hot vector with a length of 5 with the i-th element being 1 and al...

  6. [6]

    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,” in Proc. Inter- speech, 2021, pp. 2127–2131

  7. [7]

    Comparision with baseline methods We first compare the proposed SAMOS with the baselines

    Results and Analysis 4.1. Comparision with baseline methods We first compare the proposed SAMOS with the baselines. As shown in Table 1, the experimental results on the BVCC dataset indicated that the proposed SAMOS significantly outperformed baseline models on three system-level metrics emphasizing cor- relation. Compared to SSL-MOS, it also reduced syst...

  8. [8]

    We can see that removing the semantic module resulted in the degradation of all the metrics on both datasets, indicating the importance of semantic repre- sentations from SSL model

    We first investigate the contribution of the semantic infor- mation, acoustic information, and listener ID informantion to the overall model performance. We can see that removing the semantic module resulted in the degradation of all the metrics on both datasets, indicating the importance of semantic repre- sentations from SSL model. When we removed the a...

Show all 41 references
  1. [9]

    To improve prediction accuracy, SAMOS employs parallel regression and classification heads, and finally outputs the final MOS score through an aggregation layer

    Conclusions This paper presents a novel MOS prediction model called SAMOS which simultaneously utilizes semantic and acoustic information as input. To improve prediction accuracy, SAMOS employs parallel regression and classification heads, and finally outputs the final MOS sco...

  2. [10]

    Mel-cepstral distance measure for objective speech quality assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech quality assessment,” in Proc. PACRIM, 1993, pp. 125–128

  3. [11]

    considered listener information and added the scores from each listener as input, achieving some progress in prediction accuracy. Recently, with the rise of self-supervised-learning (SSL) based models trained on large-scale unlabeled data, fine- tuning SSL models and extractin...

  4. [12]

    SDR– half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR– half-baked or well done?” in Proc. ICASSP, 2019, pp. 626–630

  5. [13]

    Per- ceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Per- ceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,” in Proc. ICASSP, 2001, pp. 749–752

  6. [14]

    A short- time objective intelligibility measure for time-frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “A short- time objective intelligibility measure for time-frequency weighted noisy speech,” in Proc. ICASSP, 2010, pp. 4214–4217

  7. [15]

    ViSQOL: An objective speech quality model,

    A. Hines, J. Skoglund, A. C. Kokaram, and N. Harte, “ViSQOL: An objective speech quality model,” EURASIP Journal on Audio, Speech, and Music Processing, vol. 2015, pp. 1–18, 2015

  8. [16]

    AutoMOS: Learning a non- intrusive assessor of naturalness-of-speech,

    B. Patton, Y . Agiomyrgiannakis, M. Terry, K. Wilson, R. A. Saurous, and D. Sculley, “AutoMOS: Learning a non- intrusive assessor of naturalness-of-speech,” arXiv preprint arXiv:1611.09207, 2016

  9. [17]

    Quality-Net: An end-to-end non-intrusive speech quality assessment model based on blstm,

    S.-w. Fu, T. Yu, H.-T. Hwang, and H.-M. Wang, “Quality-Net: An end-to-end non-intrusive speech quality assessment model based on blstm,” in Proc. Interspeech, 2018, pp. 1873–1877

  10. [18]

    MOSNet: Deep learning-based objec- tive 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 objec- tive assessment for voice conversion,” inProc. Interspeech, 2019, pp. 1541–1545

  11. [19]

    MBNet: MOS prediction for synthesized speech with mean-bias network,

    Y . Leng, X. Tan, S. Zhao, F. Soong, X.-Y . Li, and T. Qin, “MBNet: MOS prediction for synthesized speech with mean-bias network,” in Proc. ICASSP, 2021, pp. 391–395

  12. [20]

    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,” in Proc. ICASSP, 2022, pp. 896–900

  13. [21]

    Generaliza- tion ability of mos prediction networks,

    E. Cooper, W.-C. Huang, T. Toda, and J. Yamagishi, “Generaliza- tion ability of mos prediction networks,” in Proc. ICASSP, 2022, pp. 8442–8446

  14. [22]

    Deep learning-based non-intrusive multi- objective speech assessment model with cross-domain features,

    R. E. Zezario, S.-W. Fu, F. Chen, C.-S. Fuh, H.-M. Wang, and Y . Tsao, “Deep learning-based non-intrusive multi- objective speech assessment model with cross-domain features,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 54–70, 2022

  15. [23]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 3451–3460, 2021

  16. [24]

    The V oiceMOS Challenge 2022,

    W.-C. Huang, E. Cooper, Y . Tsao, H.-M. Wang, T. Toda, and J. Ya- magishi, “The V oiceMOS Challenge 2022,” inProc. Interspeech, 2022, pp. 4536–4540

  17. [25]

    A transfer and multi-task learning based approach for MOS predic- tion,

    X. Tian, K. Fu, S. Gao, Y . Gu, K. Wang, W. Li, and Z. Ma, “A transfer and multi-task learning based approach for MOS predic- tion,” in Proc. Interspeech, 2022, pp. 5438–5442

  18. [26]

    DDOS: A MOS predic- tion framework utilizing domain adaptive pre-training and distri- bution of opinion scores,

    W.-C. Tseng, W.-T. Kao, and H.-y. Lee, “DDOS: A MOS predic- tion framework utilizing domain adaptive pre-training and distri- bution of opinion scores,” in Proc. Interspeech, 2022, pp. 4541– 4545

  19. [27]

    UTMOS: UTokyo-SaruLab system for V oice- MOS Challenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: UTokyo-SaruLab system for V oice- MOS Challenge 2022,” in Proc. Interspeech, 2022, pp. 4521– 4525

  20. [28]

    Fusion of self-supervised learned models for MOS pre- diction,

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

  21. [29]

    Ensem- ble of deep neural network models for MOS prediction,

    M. Kune ˇsov´a, J. Matou ˇsek, J. Lehe ˇcka, J. ˇSvec, J. Mich ´alek, D. Tihelka, M. Bul´ın, Z. Hanzl´ıˇcek, and M. ˇRez´aˇckov´a, “Ensem- ble of deep neural network models for MOS prediction,” in Proc. ICASSP, 2023, pp. 1–5

  22. [30]

    RAMP: Retrieval- augmented MOS prediction via confidence-based dynamic weighting,

    H. Wang, S. Zhao, X. Zheng, and Y . Qin, “RAMP: Retrieval- augmented MOS prediction via confidence-based dynamic weighting,” in Proc. Interspeech, 2023, pp. 1095–1099

  23. [31]

    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,” in Proc. ICASSP, 2023, pp. 1–5

  24. [32]

    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

  25. [33]

    BiV ocoder: A bidirec- tional neural vocoder integrating feature extraction and waveform generation,

    H.-P. Du, Y .-X. Lu, Y . Ai, and Z.-H. Ling, “BiV ocoder: A bidirec- tional neural vocoder integrating feature extraction and waveform generation,” arXiv preprint arXiv:2406.02162, 2024

  26. [34]

    SQAT-LD: Speech quality assessment transformer utilizing listener depen- dent modeling for zero-shot out-of-domain MOS prediction,

    K. Shen, D. Yan, L. Dong, Y . Ren, X. Wu, and J. Hu, “SQAT-LD: Speech quality assessment transformer utilizing listener depen- dent modeling for zero-shot out-of-domain MOS prediction,” in Proc. ASRU, 2023, pp. 1–6

  27. [35]

    How do voices from past speech synthesis challenges compare today?

    E. Cooper and J. Yamagishi, “How do voices from past speech synthesis challenges compare today?” arXiv preprint arXiv:2105.02373, 2021

  28. [36]

    The Blizzard Challenge 2019,

    Z. Wu, Z. Xie, and S. King, “The Blizzard Challenge 2019,” in Proc. Blizzard Challenge Workshop, 2019

  29. [37]

    Conformer: Convolution- augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu et al. , “Conformer: Convolution- augmented transformer for speech recognition,” in Proc. Inter- speech, 2020, pp. 5036–5040

  30. [38]

    ConvNeXt V2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “ConvNeXt V2: Co-designing and scaling convnets with masked autoencoders,” in Proc. CVPR, 2023, pp. 16 133–16 142

  31. [39]

    Improving self-supervised learning-based MOS prediction networks,

    B. Gyires-T ´oth and C. Zaink ´o, “Improving self-supervised learning-based MOS prediction networks,” arXiv preprint arXiv:2204.11030, 2022

  32. [40]

    ESP- Net: End-to-end speech processing toolkit,

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. E. Y . Soplin, J. Heymann, M. Wiesner, N. Chen et al., “ESP- Net: End-to-end speech processing toolkit,” in Proc. Interspeech, 2018, pp. 2207–2211

  33. [41]

    CSTR VCTK cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, K. MacDonald et al., “CSTR VCTK cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” University of Edinburgh. The Centre for Speech Technology Research (CSTR), 2019

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.