SCOREQ: Speech Quality Assessment with Contrastive Regression

In this paper, we present SCOREQ, a novel approach for speech quality prediction. SCOREQ is a triplet loss function for contrastive regression that addresses the domain generalisation shortcoming exhibited by state of the art no-reference speech quality metrics. In the paper we: (i) illustrate the problem of L2 loss training failing at capturing the continuous nature of the mean opinion score (MOS) labels; (ii) demonstrate the lack of generalisation through a benchmarking evaluation across several speech domains; (iii) outline our approach and explore the impact of the architectural design decisions through incremental evaluation; (iv) evaluate the final model against state of the art models for a wide variety of data and domains. The results show that the lack of generalisation observed in state of the art speech quality metrics is addressed by SCOREQ. We conclude that using a triplet loss function for contrastive regression improves generalisation for speech quality prediction models but also has potential utility across a wide range of applications using regression-based predictive models.

Paper

References (68)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer NTty6/10 · confidence 4/52024-07-02

Summary

The paper proposed a reference-free speech quality prediction framework. This work is based on NOMAD and uses triplet loss for contrastive regression to mitigate the generalization problem of reference-free speech quality metrics. The author conduct experiments on various datasets to show the L2 loss failed to generalize and the proposed method works.

Strengths

● The work uses triplet loss to mitigate the generalization problem of reference-free speech quality metrics. ● Experiments are done with open datasets and achieve notable improvements in generalization. ● The experiments are comprehensive. The results demonstrate the advantages of triplet loss over L2 loss.

Weaknesses

● In lines 165-167, the authors state "Our experiments show that the adaptive margin’s contribution is minimal, while significant improvement is achieved through our batch-all strategy", but there is no ablation study about adaptive margin and batch-all strategy. ● Some spelling mistakes: Figure1 "while SCOREQ quality"; Line 123 "and encoder";

Questions

● In the attached codes, the authors use the original wav2vec2 base checkpoint. But in Section 3, the paper mentioned the w2v model is finetuned with SCOREQ loss or trained from scratch, is it a model not mentioned in the paper? ● Why the Adapt variant in Table 11 is much lower than the Const variant?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See weaknesses

Reviewer FFmN6/10 · confidence 3/52024-07-12

Summary

The paper proposes a triplet loss function for contrastive regression for MOS prediction, which helps the model learn more about the relative rank of speeches.

Strengths

1. The author proposes to solve an important problem in the field of MOS prediction that the L2_loss lacks the awareness of rank. 2. The proposed loss function encourage the model to learn more about rank. 3. Experiments verifies the effectiveness of the proposed method.

Weaknesses

See the questions.

Questions

1. Why the distance is the absolute values. If the author is 2, then the pair of (0.5, 4) is the same as (3.5, 4) since the distance is both (1.5^2, 2^2)? In this way, how can we know that the quality of a sample is better or worse than the anchor? 2. Table 1 is quite important, whose explanation should be included in the main context to support the motivation of the paper. Moreover, the Appendix H is not clear enough. Can author explain it in more details? 3. It will be much easier for reader to understand if the "SCOREQ Training" section can have a figure and more detailed discuss the two mode (NR; NMR). 4. For the NMR mode, how to calculate the PC since the distance is the absolute values?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Reviewer XmpT8/10 · confidence 5/52024-07-14

Summary

The paper presents SCOREQ, a novel approach for speech quality prediction using a triplet loss function for contrastive regression. SCOREQ addresses domain generalization issues in current no-reference speech quality metrics by incorporating Mean Opinion Score (MOS) labels and improving training efficiency. Unlike NOMAD, which uses the Neurogram Similarity Index Measure (NSIM) and offline triplet preparation, SCOREQ uses in-batch triplet combinations and MOS for better alignment with human perception. The results show that SCOREQ significantly improves generalization and robustness over existing methods, making it a more effective solution for speech quality assessment.

Strengths

The paper exhibits several notable strengths across originality, quality, clarity, and significance: Originality: The paper presents a novel approach for speech quality prediction by introducing SCOREQ, which leverages a triplet loss function for contrastive regression. This approach innovatively addresses the shortcomings of existing no-reference speech quality metrics, particularly through the use of MOS labels to align better with human perception. Quality: The experimental results are extensive and robust, showcasing thorough benchmarking and comparison with state-of-the-art methods. The authors provide detailed insights into the architectural design decisions and their impact on performance, demonstrating a deep understanding of the problem domain. Clarity: The distinction between SCOREQ and NOMAD is clearly articulated, highlighting the improvements in training efficiency and performance. The paper effectively communicates the rationale behind replacing NSIM with MOS and the benefits of using in-batch triplet combinations. Significance: The work addresses the important problem of speech quality estimation, a critical area for numerous applications in speech processing. The results show high correlation with human perception, underscoring the practical relevance and potential impact of SCOREQ. Additionally, the paper demonstrates the utility of the proposed approach in both no-reference and non-matching-reference scenarios, supported by comprehensive ablation studies that validate the effectiveness of the proposed method.

Weaknesses

The paper presents several weaknesses that could be addressed to improve its overall impact and effectiveness (as also mentioned in the paper itself): Limited Test Datasets: The evaluation is primarily focused on certain domains of speech quality, lacking diversity in test datasets. Specifically, domains such as voice conversion and others are underrepresented. Broadening the scope of datasets could strengthen the generalization claims and applicability across varied speech tasks. Scope of Application: While the SCOREQ loss shows promise for various regression tasks, the study is limited to the speech domain. Exploring and evaluating the loss function in other regression contexts would provide a more comprehensive understanding of its potential and versatility. The current limitation narrows the impact of the proposed method. Memory Requirements: The SCOREQ loss function requires three times the memory compared to the L2 loss, due to the need for backpropagation on all three samples in the triplet. This could be a significant drawback for practical implementation, especially in memory-constrained environments. Strategies to mitigate this memory overhead or discussions on feasible solutions would enhance the practicality of the method. Experimental Scope: The experiments, while extensive, could benefit from additional comparisons with a broader range of state-of-the-art models. Including more baseline models and alternative approaches in the experimental setup would provide a more robust validation of the proposed method's superiority.

Questions

1. Correlation with Human Judgments: Could you provide more detailed analysis on how well the SCOREQ framework correlates with human judgments on both a sample and distribution level? Specifically, it would be valuable to see if the framework can capture variances in human perception on a population level. This additional data would strengthen the claims of alignment with human subjective assessments. 2. Evaluation on Diverse Domains: Have you considered evaluating SCOREQ on more diverse datasets, such as those from different speech tasks like voice conversion? Expanding the range of test datasets could provide more insights into the generalizability of your method. 3. Application Beyond Speech Domain: While your study focuses on speech quality assessment, do you have any preliminary insights or plans for applying SCOREQ to other regression tasks beyond the speech domain? Exploring this could help in understanding the broader applicability of your proposed loss function.

Rating

8

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

NA

Reviewer Sj6A6/10 · confidence 4/52024-07-16

Summary

This paper proposes a novel speech quality assessment method based on contrastive regression. Many speech quality assessment predictors are obtained based on supervised training by estimating an MOS value given an audio. However, it is well-known that these supervised training methods have serious over-fitting issues with the trained speech data domain and cannot be used for the other domains (e.g., DNSMOS trained with noisy-clean speech data cannot be used for TTS quality assessment and vice versa). To tackle the problem, the paper proposes a novel contrastive regression method, which extends NOMAD with several aspects (e.g., dynamic training, contrastive regression, etc.). The experimental results clearly show the improvement from NOMAD and the robustness of the unseen domain.

Strengths

- As the authors said, speech quality assessment would become an important research topic as generative speech AI studies become active. - Reasonable extensions from prior studies (NOMAD) - Showing the robust performance across the domain

Weaknesses

- The paper has several readability issues. The introduction generally uses some technical terms without explanations (or needs more understanding of the background techniques). For example, - The explanations (especially the introduction part) are twisted. Section 1 uses Figure 4 as an example, but it is very difficult to understand since we do not have explanations about the related terminologies and backgrounds in Figure 4—Ditto for line 38 about Figure 1 (a). - "Rank-and-Contrast (RnC)" suddenly appears without any explanations. - The notation is not consistent. ${\boldsymbol x}$ and $x$ are mixed. - Missing references. There have been several attempts to resolve the domain mismatch issue in speech quality assessment prediction by using self-supervised learning or unsupervised methods, e.g., [1] and [2]. I recommend the authors explore this research direction and compare them (I'm sure the method can outperform this because of the use of more supervised information compared with them). [1] Fu, Szu-Wei, et al. "Self-Supervised Speech Quality Estimation and Enhancement Using Only Clean Speech." The Twelfth International Conference on Learning Representations. [2] Maiti, Soumi, et al. "Speechlmscore: Evaluating speech generation using a speech language model." ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023.

Questions

Can this approach be applied to general audio and music generation applications? There are a lot of demands on evaluating the quality of them.

Rating

6

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

The paper describes the limitations of the proposed method in terms of scope, memory usage, and a cross-domain issue, well. It also describes the societal impact of generative model research in speech.

Reviewer 2CaS6/10 · confidence 4/52024-07-27

Summary

The paper "SCOREQ: Speech Quality Assessment with Contrastive Regression" tackles the challenges of Out-of-Distribution (ODS) and Out-of-Domain (ODM) problems by introducing a novel approach using a triplet loss function designed for contrastive regression. ### Handling ODS and ODM Problems (Domain Generalization Challenges) State-of-the-art no-reference speech quality models often fail to generalize well to unseen audio degradations. This domain mismatch is particularly problematic given the rapid advancements in generative speech technologies such as neural speech coding, speech enhancement, and speech synthesis. The authors highlight that existing methods struggle with mapping high-dimensional speech data to a low-dimensional quality space, leading to poor generalization across different domains. ### SCOREQ Approach Contrastive Regression: SCOREQ employs a triplet loss function designed for contrastive regression to create a quality manifold. This manifold is structured to order embeddings based on their similarity to MOS, thereby improving the model's ability to generalize to new domains. - On-the-Fly Triplet Mining: Unlike traditional methods that use a fixed set of triplets, SCOREQ generates triplets dynamically during training. This ensures that the model is exposed to a diverse set of triplets, enhancing its robustness to different types of audio degradations. - Adaptive Margin: The main novelty of this work is the incorporation of an adaptive margin in its loss function, which adjusts based on the MOS distance between samples. This margin helps the model learn a more nuanced representation of quality differences, further improving generalization. ### Batch-All for Regression Technique The Batch-All strategy for regression is adapted from person re-identification tasks but modified to suit the continuous nature of MOS labels. This technique ensures that every possible triplet combination within a batch is considered during training, promoting a more comprehensive learning process. - Triplet Loss for Regression: The triplet loss used in SCOREQ is formulated to ensure that the embedding distance between an anchor and a positive sample (closer in quality) is smaller than the distance between the anchor and a negative sample (further in quality). - Mask Generation: For each batch, a 3D mask is generated to identify valid triplets. This mask ensures that only triplets where the MOS distance between the anchor and the positive is smaller than the MOS distance between the anchor and the negative are used in training. - Loss Function: The SCOREQ loss function ensures that valid triplets contribute to the learning process, promoting the creation of a quality-aware embedding space. - Adaptive Margin: The adaptive margin, a **key innovation** of this work, replaces the fixed margin with one that dynamically adjusts based on the MOS distances. This allows the training to adapt more effectively to the continuous nature of speech quality, ensuring that the model learns more accurate and generalizable representations. ### Summary The SCOREQ model leverages a contrastive regression framework with a batch-all triplet mining strategy and an adaptive margin to improve the generalization of speech quality metrics. By ensuring that the embeddings are ordered based on quality differences and dynamically generating valid triplets during training, SCOREQ addresses the domain generalization issues, making it robust against both ODS and ODM scenarios. The adaptive margin, in particular, stands out as the main novelty of this work, significantly enhancing the model's ability to generalize across diverse and unseen speech quality conditions.

Strengths

The paper presents a novel approach to speech quality assessment by introducing a contrastive regression framework, SCOREQ, which employs a triplet loss function adapted for regression tasks. This approach is original in several ways: - Adaptive Margin: The main novelty lies in the adaptive margin, which adjusts dynamically based on the MOS distances between samples. This innovative use of an adaptive margin for regression tasks enhances the learning process by ensuring more accurate and nuanced representation of quality differences. - Contrastive Regression: While contrastive learning has been extensively used in classification tasks, its application to regression tasks, particularly in speech quality assessment, is novel. The paper successfully adapts and extends the principles of contrastive learning to handle continuous labels, addressing the domain generalization shortcomings of existing metrics. - On-the-Fly Triplet Mining: The dynamic generation of triplets during training is a creative solution to the limitations of fixed triplet sets, providing a more comprehensive and robust learning experience for the model. Quality The significance of the paper is substantial: - Improved Generalization: By addressing the domain mismatch issues in speech quality assessment, SCOREQ provides a significant advancement in the field. Its ability to generalize well across diverse and unseen conditions makes it highly valuable for real-world applications. - Wide Applicability: The principles and techniques introduced in the paper, such as contrastive regression with adaptive margins, have potential applications beyond speech quality assessment. *They could be adapted for other regression-based tasks in various domains*. Practical Impact: The ready-to-use speech quality metrics developed from SCOREQ can be directly applied in various speech processing applications, such as neural speech coding, speech enhancement, and synthesis. This practical impact underscores the importance of the contributions made by the paper.

Weaknesses

- While the paper demonstrates strong performance of SCOREQ in speech quality assessment, its evaluation is limited to this specific domain. The authors suggest that the method could be applicable to other regression-based tasks, but they do not provide empirical evidence to support this claim. To strengthen the paper, the authors could discuss more thoroughly the potential applications of SCOREQ in other domains and possibly provide preliminary results or case studies to illustrate its broader applicability. - The adaptive margin and batch-all strategy, while innovative, introduce additional computational complexity and memory usage. The paper briefly mentions this but does not provide a detailed analysis of the computational overhead and how it compares to existing methods. Including a more detailed discussion on the computational costs and providing quantitative comparisons of training times and memory usage with other state-of-the-art methods would help practitioners understand the trade-offs involved. - The paper does not address the potential for real-time applications of SCOREQ. Real-time processing is critical for many practical applications of speech quality assessment. The authors could provide insights into the feasibility of real-time implementations of SCOREQ, possibly suggesting optimizations or trade-offs that could make real-time usage more viable.

Questions

- It is mentioned that the SCOREQ method could be applicable to other regression-based tasks. Can you provide more details or preliminary results on how SCOREQ performs in other domains? - How does the computational complexity of SCOREQ compare to traditional L2 loss models and other state-of-the-art methods in terms of training time and memory usage? - While your evaluation covers multiple datasets, do you believe these datasets comprehensively cover the variability in speech quality assessment? What are the potential limitations of your chosen datasets? - Have you considered the potential for real-time applications of SCOREQ? If so, what are the main challenges and how might they be addressed? - Can you elaborate on the rationale behind your choice of hyperparameters, particularly for the adaptive margin and triplet loss configurations? How sensitive is SCOREQ to these hyperparameters?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have made an effort to discuss the limitations of their work, particularly in the section on computational complexity and domain-specific evaluation. However, there are areas where the discussion could be more thorough: 1. Memory and Computational Requirements: The paper briefly mentions that the SCOREQ loss requires three times the memory of the L2 loss due to the need to compute the forward pass for anchor, positive, and negative samples. A more detailed analysis of the computational and memory requirements, including quantitative comparisons with other methods, would provide a clearer understanding of the trade-offs involved. This should include discussions on the feasibility of deploying SCOREQ on different hardware setups, especially those with limited resources. 2. Domain-Specific Evaluation: The evaluation of SCOREQ is primarily limited to speech quality assessment. While this domain is well-covered, the broader applicability of SCOREQ to other regression tasks is not empirically validated. The authors should acknowledge this limitation more explicitly and propose future work to explore and validate the generalizability of SCOREQ in other domains. This could include preliminary results or theoretical discussions on potential applications outside speech quality assessment. 3. Real-Time Application Feasibility: The potential for real-time applications is not addressed in the paper. The authors should discuss the challenges and feasibility of implementing SCOREQ for real-time applications. This includes potential optimizations and trade-offs that might be necessary to achieve real-time performance.

Reviewer Sj6A2024-08-08

I appreciate the authors' additional experiments, but as the authors say, this is not a fair comparison, and it would not change my overall impressions. Also, I still feel that the explanations are twisted (thanks for the detailed explanations, but these explanations repeat the current paper structure). I think it is better to have brief explanations about them and may also add some remarks that they are explained later in Section 4 if needed. I could not expect significant improvement in the effectiveness and readability of this paper. Thus, I want to keep my score as it is. > Thank you for pointing out this potential inconsistency. We have carefully reviewed our notation throughout the paper and believe we have consistently used boldface to denote vectors (e.g., x) and non-boldface to denote scalars (e.g., x). However, we are open to making any necessary corrections. To assist us in doing so, could you please specify the line number(s) where you believe the notation is inconsistent? We will gladly address any specific instances you identify. Do you mean that $x_a$, $x_p$, and $x_n$ lines 133 and $x_i$ and $x_{test}, x_{nmr}$ in line 178--182 are scalar? If so, please explain them. Also, subscripts and superscripts are mixed ($a$, $p$, $n$, $i$ appear as subscripts and superscripts in equations 1--5 and inline explanations in page 4.). If they are intentional, please explain it.

Reviewer FFmN2024-08-11

Thanks for your response

Thanks for the detailed response, which has solved most of my concerns. I have adjusted my score accordingly.

Reviewer XmpT2024-08-13

Response to rebuttal

I would like to thank the authors for their rebuttal. I have read the reviews and rebuttals and at the moment would like to keep my score the same.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC