We sincerely thank Reviewer XEXh for the thoughtful, constructive, and detailed feedback. Your comments have been instrumental in helping us clarify technical aspects, highlight the strengths of ARECHO, and identify areas for improvement. We deeply appreciate your time and effort in reviewing our work. Below, we carefully address each point raised and outline corresponding clarifications and revisions.

Q1: Clarification on Quantization Details

We appreciate your request to clarify our quantization strategy. In response, we will revise Appendix A to include concrete illustrations of our percentile-based tokenization and its impact on model behavior.

Rather than performing an exhaustive search per metric, we use a consistent percentile-based tokenization strategy with 500 bins for all continuous-valued metrics. This enables uniform treatment across diverse metrics and simplifies model design. In Table 10, we conducted an ablation study on bin size, showing that 500 bins present a good tradeoff between reconstruction quality and ARECHO prediction difficulty. More importantly, we would like to reinforce why this design matters. ARECHO’s tokenization strategy unifies both regression and classification metrics under a single modeling framework, this is a key architectural innovation. Unlike UniVERSA, which relies on direct regression, our token-based method ensures consistent modeling across heterogeneous metrics, paving the way for robust scaling and modular extension.

While ordinal regression is indeed a viable alternative, we deliberately opted for token-based quantization to support generalization, controllable sampling, and eventual integration with language-driven generation. We fully agree this is a rich area for future exploration, and will add a discussion on hybrid modeling as a natural next step in evolving ARECHO.

We will expand Appendix A with visualizations and deeper explanation of this strategy, and we will highlight this design motivation in Section 3.2.

Q2: Order Effect on Metric Prediction

Thank you for raising the important question regarding the ordering of metrics during inference. We appreciate the opportunity to clarify that ARECHO supports two inference modes:

• In the fixed order mode, users manually define the sequence of metrics to be queried.
• In the flexible mode, our proposed two-step confidence-oriented decoding enables the model to dynamically determine the most effective order.

This flexibility is not an afterthought, it is central to ARECHO’s ability to adapt to incomplete annotations and metric sparsity. Our results in Appendix J and K demonstrate that flexible decoding not only preserves performance but also improves interpretability by aligning prediction order with metric dependencies. This makes ARECHO uniquely positioned among multi-metric estimators to handle real-world uncertainty in deployment scenarios.

We will revise the text to more clearly present this distinction between training and inference order and explain the robustness benefits enabled by dynamic ordering.

Q3.1: Importance of Inter-metric Relationships

We strongly agree that inter-metric relationships are crucial, and ARECHO places them at the heart of its design. Unlike UniVERSA and TorchSquim, which treat metrics independently, ARECHO models their semantic and statistical dependencies directly through chain-based autoregression.

This leads to two significant benefits:

• Performance gains across all major metrics and datasets (Table 1),
• Interpretability, as demonstrated in our dependency analysis (Appendix J).

We emphasize that ARECHO is not limited by this coupling: when needed, it can revert to single-metric inference, operating efficiently and independently. This optionality makes ARECHO not only accurate but controllable and practical in real-world settings.

Q3.2: Practical Usage of ARECHO

We appreciate the reviewer’s interest in the real-world applicability of ARECHO. As discussed in Section 5, ARECHO is applicable to a wide range of scenarios, including:

• Evaluation of corrupted speech signals
• Text-to-speech (TTS) systems
• Speech enhancement pipelines

We thank you for highlighting the opportunity to make this clearer in the paper. We will revise the section to better emphasize this versatility.

More importantly, ARECHO is ready to be used beyond evaluation. It can serve as:

• A reward model for reinforcement learning setups
• A component in multi-metric loss functions during training of other systems

These are exciting directions that we aim to explore in future work. We would like to provide some references where previous studies utilize different metrics to enhance their speech generation systems (e.g., TTS systems or enhancement systems)

• Preference alignment improves language model-based TTS (Tian et al. 2025)
• Aligning Generative Speech Enhancement with Human Preferences via Direct Preference Optimization (Li et al. 2025)
• Multi-CMGAN+/+: Leveraging multi-objective speech quality metric prediction for speech enhancement (Close et al. 2024)
• Improving Speech Enhancement with Multi-Metric Supervision from Learned Quality Assessment (Wang et al. 2025)

Lastly, we would like to highlight that ARECHO achieves over 100× speedup compared to computing ground-truth metrics. This efficiency is critical for real-time and large-scale deployment. To better reflect this advantage, we will revise the manuscript to include a detailed comparison between ARECHO and ground-truth metric computation, which is currently underexplored in the text.

Q3.3: Usage for Long Audio

We thank the reviewer for the thoughtful suggestion to consider long-form audio. We fully agree that temporal dependency modeling across extended audio contexts is an important next step. Accordingly, we will explicitly include this point as a direction for future work in the revised manuscript.

Q4: Analysis on MOS Data Evaluation

We sincerely appreciate the reviewer’s suggestion to conduct deeper analysis on MOS metrics. In response, we conducted a detailed evaluation to examine how inter-metric relationships impact the modeling of MOS-related metrics. As shown below, ARECHO generally outperforms UniVERSA-T across multiple datasets.

Main metrics comparison (demonstrating benefits of inter-metric prediction):

DNSMOS ablation (ARECHO-full: multi-metric prediction, ARECHO-only: single-metric prediction):

We thank the reviewer again for prompting this analysis. Due to space constraints, we will include extended results and deeper analysis in the revised manuscript and appendix.

Final Remarks

We sincerely thank Reviewer XEXh once again for their constructive and detailed feedback. Your questions helped us clarify architectural motivations, improve empirical presentation, and strengthen the practical framing of ARECHO. We are confident the rebuttal and the scheduled presentation-related revision will address your concerns and more clearly convey the significance and applicability of our contributions.

We sincerely thank Reviewer z2T7 for their thoughtful and constructive feedback. Your comments helped us further clarify the motivation and scope of our work, and we greatly appreciate the opportunity to address them in detail. Below, we respond point by point and outline the revisions we will make to strengthen the manuscript.

W1: Ordinal Regression Loss

Our core focus in this work is to develop a unified framework capable of modeling both regression and categorical metrics via a tokenized sequence prediction paradigm. This tokenization strategy enables consistent modeling of heterogeneous evaluation targets and facilitates scalable autoregressive inference.

We fully acknoweldge that many metrics, such as MOS and SNR, are ordinal in nature and that modeling them using ordinal-specific surrogate losses is likely to improve performance and fidelity. We appreciate the reviewer’s references to SLACE and adversarial ordinal loss as concrete and compelling alternatives. While our current formulation focuses on a unified classification-based modeling pipeline to maintain consistency across all metric types, we view integrating ordinal-aware objectives as a high-impact direction for enhancing metric-specific accuracy. We will explicitly incorporate this direction in our Future Work section, acknowledge the previous works, and plan follow-up experiments to validate this hybrid design.

W2 & Q4: Use of Macro-averaged MAE

We are grateful for the emphasis on macro-averaged MAE. This perspective aligns with our intention to equitably reflect performance across frequent and rare classes, particularly given the diversity of our evaluation metrics. We clarify that our current evaluation already follows the macro-averaged MAE formulation as discussed in [Gutiérrez et al., 2016], but we acknowledge the need to make this design choice more visible.

We will update both the main text and Appendix D to clearly state this design choice and its rationale.

Q1: Effect of Beam Size B

We thank the reviewer for raising this clarification. The impact of beam size B is empirically evaluated in Ablation I.2, where we demonstrate that a small beam size (B=2) achieves an effective trade-off between prediction accuracy and decoding efficiency.

To improve accessibility, we will revise Appendix D to include a direct pointer to this analysis and briefly summarize the key findings to guide readers.

Q2: Subjective Metric Evaluation

We agree with the reviewer that evaluating subjective metrics is essential for understanding the practical utility of the system. In response, we have expanded our experimental results to include detailed evaluations on subjective metrics such as NISQARealMOS and URGENTMOS.

These results demonstrate the metric-specific nature of performance. For NISQARealMOS, which captures perceptual quality, our proposed autoregressive model (ARECHO base) achieves substantial improvements across all metrics, MAE: 0.12, RMSE: 0.23, SRCC: 0.98, and Kendall’s Tau: 0.92, compared to the non-token UniVERSA base baseline (MAE: 0.85, SRCC: 0.81) and token baseline UniVERSA-T base (MAE: 0.30, SRCC: 0.94). This highlights strong alignment between the predicted and ground-truth perceptual scores.

In contrast, for URGENT_MOS, which reflects speech enhancement subjective evaluation, the improvements are less pronounced. While UniVERSA-T base performs best (MAE: 0.33, SRCC: 0.92), the ARECHO model shows moderate gains over the UniVERSA base baseline (MAE: 0.41 vs. 0.58, SRCC: 0.71 vs. 0.67), indicating room for further improvement on certain subjective metrics.

We will report both ranking-based (e.g., SRCC, Kendall’s Tau) and error-based (e.g., MAE, RMSE) metrics, and include the full results in Appendix L, with key summary findings highlighted in Section 5.

Q3: Use of Ground-Truth Values During Inference

We appreciate the reviewer’s thoughtful question regarding the role of ground-truth values in inference. ARECHO is specifically designed for efficient and generalizable inference, especially in settings where ground-truth computation is impractical, such as real-time evaluation, black-box systems, reference-dependent metrics, or large-scale deployment.

Our experiments show that ARECHO achieves over 100× speedup compared to computing all metrics using their original estimators (e.g., VERSA). This efficiency is critical for applications requiring rapid and scalable evaluation.

Injecting ground-truth at inference would limit applicability, but we recognize that hybrid decoding strategies may promote a useful trade-off between performance and realism. We will include this hybrid strategy as a promising future direction and add a brief feasibility discussion in Appendix L.

Limitations: Potential Societal Impacts

We thank the reviewer for encouraging reflection on broader implications. While automated multi-metric estimators can substantially accelerate system development and iteration, they should not be viewed as replacements for human listening tests, especially in high-stakes or perceptually nuanced applications. Final evaluation should always include human oversight to ensure alignment with subjective experience.

We acknowledge the potential societal risks that may arise if these automated predictions are misinterpreted without appropriate context or used in sensitive domains such as healthcare or biometric identification. To mitigate such risks, we advocate for: transparent model design, human-in-the-loop evaluation, and restricted deployment protocols. We will include this discussion in the main manuscript ("limitation section") to promote responsible development and deployment.

Final Remarks

We are deeply grateful for Reviewer z2T7’s valuable feedback. The issues you raised have significantly improved the clarity, robustness, and broader framing of our work. At the same time, we are also excited to see that the proposed ARECHO framework can serve as a promising anchor for multiple future directions, including integrating ordinal regression losses, exploring hybrid models that incorporate ground-truth supervision, and enabling practical uses such as reward modeling for speech generation and foundational evaluation in broader applications.

We sincerely thank Reviewer DC5A for the detailed and constructive feedback. Your comments helped us strengthen the clarity, technical analysis, and experimental validation of our work. Below, we address each point and outline the corresponding revisions.

W1 and Q1: Measure Loss of the Quantization

We thank the reviewer for raising this important concern. Since our tokenization bins are learned from the training set, we evaluated their generalization to test data via a reconstruction analysis. Specifically, we measured how well the quantized–dequantized values match the original ground-truth metrics on the test set.

These results confirm that our quantization strategy yields low reconstruction error across domains. Notably, percentile-based tokenization significantly outperforms uniform binning, demonstrating that data-aware strategies preserve finer-grained information crucial for accurate metric prediction.

Revision: We will expand Appendix A to include these findings.

Noted that even though with 1000 bin size, the reconstruction performance is better. It cannot guarantee better performance in ARECHO as it also introduce more difficult prediction targets. Practically, we find 500 bin size is a good tradeoff of effectiveness between better reconstruction and easier prediction. We have discussed the results in Table 10.

W2 & Q2: Error Accumulation and Metric Dependency

ARECHO’s autoregressive design explicitly models inter-metric dependencies, which improves both prediction quality and interpretability (as demonstrated in Table 1 and Appendix J).

To alleviate concerns about error propagation, ARECHO supports flexible inference, users can query only a subset of metrics, bypassing previously predicted values. In such cases, ARECHO operates like an independent predictor without chaining.

We here examplified experiments comparing full-chain vs. single-metric (no chaining) inference for DNSMOS (on generation test set):

DNSMOS Performance Comparison:

These results confirm that ARECHO is resilient to error accumulation while benefiting from dependency modeling when enabled.

Revision: We will clarify this flexibility in the main text and highlight these results in Appendix L.

W3: Training and Inference Complexity

ARECHO is trained using random order sampling, which is comparable in cost to standard multi-task models as evidenced in Appendix L.

Although inference is sequential, the process is highly efficient: ARECHO achieves over 100× speedup compared to computing ground-truth metrics using model-based estimators (e.g., Qwen-based metrics and models with pre-trained encoders like UTMOS, Noresqa, ScoreQ).

While parallel predictors are simpler per step, they scale poorly with large metric sets in terms of memory and compute. In contrast, ARECHO provides a scalable, accurate, and efficient alternative.

Revision: We will add a complexity analysis and comparison in Appendix L.

Q3: Confidence-Oriented vs. Greedy Decoding

We conducted an ablation comparing two-step confidence-oriented decoding versus greedy decoding on enhanced speech test set.

This supports our claim that naive greedy decoding, while simple, is suboptimal in the presence of inter-metric dependencies and dataset frequency imbalance. The confidence-based approach provides more principled, performance-driven metric ordering.

Revision: We will include this ablation in Appendix K and add a summary pointer in the main text.

We thank the reviewer again for the thoughtful suggestions, which have significantly improved both the clarity and technical depth of our work.