acceptodds
Under review as a conference paper at ICLR 2027

One Prompt, Many Voices: Benchmarking and Improving Timbre Diversity in Instruction-Following Voice Design

Abstract

Voice design requires generating natural speech that adheres to textual descriptions while providing diverse timbre candidates for applications like character casting. Since natural language descriptions constrain perceptual attributes without dictating a unique speaker identity, multiple distinct voices can satisfy a single instruction. However, current research primarily emphasizes instruction following and overlooks timbre diversity, leaving the generation of diverse yet compliant voices underexplored. To address this gap, we introduce VoxDICE, a benchmark combining hierarchical voice descriptions and a real speech reference set to jointly assess instruction following, conditional timbre variation, and global speaker distribution alignment. Evaluations of current models on VoxDICE reveal a prevalent empirical tension between instruction-following and timbre diversity. To overcome this limitation, we propose a reinforcement learning approach based on group relative policy optimization. This method integrates a distinct timbre reward within generation groups alongside an instruction gate that restricts diversity bonuses to compliant candidates. Experiments on the Qwen3 TTS model demonstrate significant improvements in both timbre diversity and instruction following, establishing that these capabilities can be jointly enhanced. Demos are available at https://voxdice744.github.io/voxdice/.

Primary area:
datasets and benchmarks
Keywords:
Voice Design, Timbre Diversity, Instruction Following

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZkfYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies an important property of voice design: a single textual specification should permit multiple distinct, instruction-compliant speaker identities. It introduces VoxDICE, with hierarchical voice descriptions and a 10,000-utterance real-speech reference set. Evaluation combines within-condition speaker-embedding diversity (MPCD and Vendi Score), global distribution matching/coverage (FSD and SRS), and instruction-following rate. Human judgments validate local diversity proxies, while a fixed-sample speaker-insertion experiment probes the global metrics. An instruction-gated, collision-based GRPO reward is applied to Qwen3-TTS, improving aggregate diversity and instruction following over its original checkpoint.

Critical strengths
  1. The benchmark defines a consequential evaluation problem and studies complementary failure modes with substantial empirical effort. Holding both the instruction and synthesis text fixed while sampling K=10 outputs is an appropriate way to expose conditional repetition. Adding global reference alignment/coverage and instruction adherence prevents any single diversity score from standing in for overall voice-design quality. The 10,000-utterance curation, three instruction scales, seven baseline systems, and GRPO intervention with gate/diversity ablations constitute meaningful benchmarking work. The reversed local/global ordering of VoxCPM2 and Ming-omni-TTS under APS illustrates why these axes are informative. Reusing Emilia and InstructTTSEval templates does not negate the contribution of constructing and validating this evaluation protocol.

  2. The paper provides substantive human validation rather than relying solely on embedding metrics or an untested LLM judge. Section 5.4 covers 195 audio groups, 1,950 pairs, and 5,850 judgments from 30 participants. Vendi Score obtains Pearson/Spearman correlations of 0.783/0.791 with perceived pairwise diversity, and MPCD obtains 0.797/0.795. Appendix A.2 separately evaluates 900 instruction-audio cases with three annotations each (2,700 judgments); human-judge agreement is 80.7%, 83.3%, and 72.3% for APS, DSD, and RP. The global-metric experiment fixes the utterance count at 10,000, varies AISHELL-2 speaker count from 20 to 1,991, and repeats sampling 20 times per count. These are material validation efforts and support the metrics as useful diagnostics, with the limits discussed below.

Critical weaknesses

Reproducibility remains a concern. Although the authors promise to release the benchmark and evaluation code upon acceptance, the currently accessible materials do not clearly support reconstructing the exact reference set and evaluation pipeline, making independent reproduction of the reported results uncertain.

The validation supports useful proxies, but does not yet establish the reliability of the benchmark's central comparative claims. The local correlations are computed on 195 groups deliberately stratified into 13 Vendi Score bins. This supports association over a broad selected range, but does not by itself establish human-consistent model rankings or discrimination between nearby systems. The paper does not report inter-annotator reliability, model/scale composition of these groups, or how the 'Unsure' option and three-way splits in the annotation interfaces are mapped to the binary majority labels. Likewise, the 900-case IFR check reports raw agreement without class balance, error breakdowns, or a human-checked comparison of the model differences central to the conclusions. The lower RP agreement warrants particular care; it does not, by itself, prove that a reported IFR difference is spurious. For global metrics, the balanced real-speech speaker-insertion test is useful but does not test robustness to the nonuniform speaker frequencies and acoustic conditions of generated speech. FSD measures reference moment alignment, while SRS extrapolates reference-entry coverage under a uniform-occupancy assumption; neither should be read as a direct count of perceptually distinct speakers. Finally, diversity is measured over all K outputs and IFR separately, leaving the intended quantity—diversity among compliant alternatives—unmeasured. RP illustrates the ambiguity: VS increases from 3.520 to 4.040 while IFR falls from 50.0% to 48.4%. These gaps limit how confidently the benchmark can rank systems for the stated use case, even though they do not invalidate the observed correlations or negate the study's workload.

Questions for authors
  1. Artifact-access and reproducibility clarification: Is an anonymous evaluation package already available to reviewers? Please identify what is accessible now versus promised after acceptance. In particular, can the exact reference IDs/segments and three-scale instructions, corpus composition and speaker-duplication audit, CAM++ checkpoint/preprocessing, scoring code, baseline configurations/seeds, and cached judge/annotation outputs be inspected or reconstructed exactly? Please also state the source-data version and reuse terms. This request concerns evaluating the benchmark; it does not require releasing the private RL training corpus.

  2. Human-validation clarification and targeted analysis: How were the 195 groups distributed across models and prompt scales, and how were 'Unsure' responses and votes with no majority handled in both human studies? Please report inter-annotator agreement and class balance/confusion information for the IFR check. Do the existing human annotations support the main model ordering, particularly the original-versus-GRPO comparison where represented, with uncertainty computed at the group/case level? A focused reanalysis of existing annotations would be more informative than simply collecting more correlated votes.

  3. Global-metric interpretation and robustness: How many distinct or conservatively clustered speakers are represented by the 10,000 reference utterances, and how does residual duplication affect the scores? Can the existing speaker-insertion experiment be reweighted to test unequal speaker frequencies and report whether the conclusions survive? Please distinguish alignment to the selected expressive-speech reference from a general claim about real-world speaker diversity, and clarify the sensitivity of the SRS extrapolation to its uniform-occupancy assumption.

  4. Targeted evidence for the benchmark's intended use: Under a fixed generation budget, can you report diversity restricted to instruction-compliant outputs together with compliant-output yield, explicitly handling cases with fewer than two valid outputs? Applying this to the main original-versus-GRPO comparison, with a human-checked subset or documented judge uncertainty, would materially affect my assessment of whether the benchmark measures diverse yet compliant voice choices.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We thank the reviewer for the careful and constructive assessment. We respond to each concern below. All section, page and line numbers refer to the revised manuscript.


W2 & Q4: Diversity among instruction-compliant outputs under a fixed budget

We agree that reporting diversity over all K outputs separately from IFR cannot show whether added diversity comes from compliant alternatives or from instruction violations. The revision measures diversity within the compliant subset in two independent ways. The first is automatic and accounts for the uncertainty of the judge. The second is a direct human count.

Valid MPCD. For each instruction and synthesis text we generate a fixed budget of K = 10 utterances. Let M be the number of utterances that the judge labels as compliant, which we call the compliant-output yield. Its expectation is K × IFR. V-MPCD is the mean pairwise cosine distance among these M utterances. When M ≤ 1, V-MPCD is set to zero and the case is kept in the average. Diversity contributed by non-compliant outputs is therefore ignored, and cases with low yield are penalized directly (Section 3.3.1, Eq. 2, page 4, lines 181–186). V-MPCD is computed on the same 1,000 cases for every model (Table 1, page 8). Models with weaker instruction following lose more diversity when moving from MPCD to V-MPCD, and RP shows the largest drop for every model. This confirms the ambiguity that the reviewer identified. For the main comparison, GRPO raises V-MPCD on every scale. This includes RP, where IFR decreases slightly, so the gain cannot be explained by additional instruction violations (Section 5.3, page 9, lines 449–452).

Judge uncertainty. V-MPCD depends on the judge's labels, so we propagate the judge's errors into it with a two-stage bootstrap (Appendix A.5, pages 19–20, lines 1008–1042).

  • In the first stage, we resample the 300 human-annotated pairs of each scale. From each resample we re-estimate two flip rates: the probability that humans reject an output the judge accepted, FP/(TP+FP), and the probability that humans accept an output the judge rejected, FN/(TN+FN).
  • In the second stage, we flip every judge label in the evaluation set according to these rates and recompute V-MPCD. Both models share the same flip rates in each of the 1,000 iterations.
ScaleV-MPCD, Qwen3-TTS → GRPOPerturbed V-MPCD, Qwen3-TTS (95% CI)Perturbed V-MPCD, GRPO (95% CI)IFR (%), Qwen3-TTS → GRPO
APS0.275 → 0.3420.290 [0.287, 0.294]0.346 [0.344, 0.349]77.9 → 83.3
DSD0.347 → 0.4170.364 [0.360, 0.369]0.411 [0.407, 0.415]81.1 → 86.6
RP0.251 → 0.2690.297 [0.285, 0.309]0.324 [0.317, 0.331]50.0 → 48.4

Judge errors shift the absolute values because flipped labels change each valid set. On every scale, however, GRPO keeps a higher V-MPCD and the 95% intervals of the two models do not overlap. This holds even on RP, where the judge agrees least with humans.

Human count of effective voices. We also ran a new study that relies on neither the judge nor the speaker encoder (Appendix A.10, Table 15, pages 23–24, lines 1231–1284).

  • For each scale, we sample 100 prompts, and each model generates five utterances per prompt.
  • Three blinded annotators first discard non-compliant utterances and then group the remaining utterances by speaker. The number of groups is the number of effective voices.
  • A distinct but non-compliant voice therefore counts as zero, and compliant voices that share a timbre count once. This is the quantity the reviewer describes: diversity among compliant alternatives under a fixed budget.
  • Uncertainty is computed at the prompt level, using a prompt-level bootstrap and paired Wilcoxon tests with Holm correction.
ScaleEffective voices, Qwen3-TTSEffective voices, GRPOHolm pHuman compliance (%), Qwen3-TTS → GRPOFleiss' κ
APS3.104 [2.990, 3.221]3.470 [3.379, 3.563]< 3×10⁻⁵76.2 → 80.60.781
DSD3.377 [3.250, 3.504]3.769 [3.654, 3.885]< 3×10⁻⁵79.4 → 84.40.762
RP2.062 [1.941, 2.188]2.277 [2.185, 2.365]1×10⁻³51.2 → 50.40.675

GRPO yields 10–12% more effective voices on every scale. On RP, the number of effective voices rises significantly while human-judged compliance is essentially unchanged, which resolves the RP ambiguity that the reviewer raised. Taken together, the automatic and human results indicate that the benchmark captures diverse yet compliant voice choices. They also indicate that GRPO increases diversity among compliant outputs, not only among all outputs.


AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

W1 & Q1: Reproducibility and artifact access

An anonymous evaluation package is now available at https://huggingface.co/datasets/ICLR45502/VoxDice_anonymous. It contains everything required to evaluate the benchmark and excludes only the private RL training corpus. We have verified that the released code reproduces our main results. The package includes:

  • Reference set. All 10,000 reference utterances are provided as the original Emilia segments, with each file named by its utterance ID, so every segment can be traced to the source data. Each utterance comes with its transcript and its APS, DSD and RP instructions.
  • Corpus audits. The corpus composition and the audit of highly similar utterances (Appendices A.2–A.3) can be reconstructed from the released audio and CAM++ embeddings.
  • CAM++. We use the public checkpoint https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common. The preprocessing code fixes the resampling, the FBank front end and the embedding normalization.
  • Scoring code. The code computes every diversity metric, including V-MPCD. It also includes the LLM-judge script with its model, prompt and decoding settings.
  • Generation code. We provide generation code for all eight systems. It uses a deterministic seed scheme: each utterance receives its own seed, derived from a fixed base seed and logged together with the output.

The cached judge outputs and human annotations are not in the repository because of their volume. They correspond one to one with the generated audio, and the released judge script can regenerate them. All reference audio comes from the Chinese subset of the original Emilia dataset, which is distributed under the CC BY-NC license, and VoxDICE follows the same non-commercial terms. Gemini-3.1-Pro-Preview generated the APS instructions and GPT-4o generated the DSD and RP instructions (Section 3.2.2, pages 3–4). The templates are given in Appendix A.14. The trained GRPO checkpoints will be released upon acceptance.


W2 & Q2: Human-consistent ranking and discrimination between nearby systems

We agree that a correlation computed on groups stratified across a wide range of Vendi Scores establishes association, but it does not show that the metrics can discriminate between nearby systems. We therefore added a two-alternative forced-choice (2AFC) study that measures how large a score difference listeners can perceive (Section 5.5, Figure 5, page 10, lines 508–527; Appendix A.12).

  • Trial design. Each trial presents two groups of ten utterances generated by two different models from the same instruction and synthesis text. The annotator selects the group with more distinct speakers, which mirrors how systems are compared in practice.
  • Tier matching. We split the case-level scores of all eight systems at their tertiles into Low, Mid and High tiers. We pair only groups that fall in the same tier under both VS and MPCD.
  • Sampling. Within each tier we sample 100 pairs from each of five |ΔVS| intervals, giving 1,500 pairs annotated by 30 participants. For MPCD, the same pairs are re-binned by |ΔMPCD|.

The table reports the win rate of the higher-scoring group as VS / MPCD.

Score difference (VS / MPCD)Low tierMid tierHigh tier
[0, 0.2) / [0, 0.02)56.0% / 55.2%48.0% / 46.9%51.0% / 53.0%
[0.2, 0.4) / [0.02, 0.04)74.0%*** / 72.2%***57.0% / 53.1%48.0% / 46.4%
[0.4, 0.6) / [0.04, 0.06)82.0%*** / 74.2%***70.0%*** / 72.5%***57.0% / 58.4%
[0.6, 0.8) / [0.06, 0.08)96.0%*** / 93.3%***81.0%*** / 82.1%***69.0%*** / 70.2%**
≥ 0.8 / ≥ 0.0898.0%*** / 98.0%***89.0%*** / 86.9%***82.0%*** / 85.0%***

** p < 0.01 and *** p < 0.001, one-sided z test against 50% with Holm correction.

These results lead to three conclusions:

  1. Rankings agree with listeners once the gap is large enough. Above a modest score gap, listeners consistently agree with the metric about which system is more diverse. In every tier, once the gap exceeds the smallest intervals, the win rate rises steadily with the gap and reaches 82–98% in the largest interval.
  2. The resolution limit for nearby systems is now explicit. No tier departs from chance in the smallest interval, so differences of that size should not be read as perceptual. The perceptibility threshold grows with the tier, from about 0.2 VS or 0.02 MPCD in the Low tier to roughly three times that in the High tier.
  3. The reported improvement is perceptible. Qwen3-TTS lies in the Low and Mid tiers. GRPO raises its average VS by 0.541 and its average MPCD by 0.061, and both gains exceed the thresholds of these tiers, so the main original-versus-GRPO ordering is audible to listeners.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

W2 & Q2: Annotation details and reanalysis of existing annotations

The revision reports all of the requested information.

  • Composition of the 195 groups. Appendix A.11 and Table 16 list the number of groups per system and scale (pages 24–25, lines 1286–1307). Groups are sampled uniformly within the VS bins, so each system's share follows its VS distribution. Every system contributes 17 to 36 groups.
  • "Unsure" and no-majority votes. Both studies use a strict rule, and no case is left undecided.
    • In the timbre study, a pair is labeled "same speaker" only when "Same" votes outnumber "Different" votes. "Unsure" counts toward neither side (Appendix A.13.1, page 26, lines 1357–1360).
    • In the IFR check, a case is labeled "follows" only when "Matches" votes outnumber "Does not match" votes (Appendix A.4, page 18, lines 961–966).
    • In both studies, all remaining cases receive the negative label.
  • Inter-annotator reliability. Fleiss' κ is 0.73 for the timbre study (Section 5.4, page 9, line 478). For the IFR check it is 0.72, 0.75 and 0.67 on APS, DSD and RP (Appendix A.4, page 18, lines 967–971). For the new effective-voice study it is 0.781, 0.762 and 0.675 (Table 15).
  • Class balance and confusion for the IFR check. Table 8 reports the full TP/FP/TN/FN counts together with the precision, recall and NPV of the judge (page 19, lines 972–1006).
    • The human-positive rate is 60.3%, 69.0% and 50.7% on APS, DSD and RP.
    • The chance-corrected judge–human agreement (Cohen's κ, computed from Table 8) is 0.59, 0.58 and 0.45.

Do the existing annotations support the original-versus-GRPO comparison? We examined this question and found that the existing annotations are too few, and not paired, to support a statistical test of this comparison.

  • In the 195-group study, Qwen3-TTS and GRPO contribute only 17 and 19 groups, spread across three scales and not matched by prompt.
  • In the IFR check, each model has 29 to 44 unpaired cases per scale. With about 40 binary labels per arm, the standard error of an IFR difference is about 9–10 percentage points, which far exceeds the 3–5 point differences at stake.

We agree with the reviewer that collecting more correlated votes would not help. We therefore ran the targeted effective-voice study described above. It is paired at the prompt level and uses 100 prompts per scale. For each prompt and model, the number of effective voices and the compliance rate are averaged over the three annotators, and the 100 prompts of each scale are treated as the independent units of analysis. The study shows with statistical significance on all three scales that GRPO produces more compliant and perceptually distinct candidate voices.


W2 & Q3: Global metrics under unequal speaker frequencies and the uniform-occupancy assumption

We reweighted the speaker-insertion experiment as the reviewer suggested (Section 3.3.3, Figure 2, page 5, lines 251–269). We sample 100 to 1,991 AISHELL-2 speakers and fix the total at 10,000 utterances. Utterances are allocated by a Zipf law with exponent s ∈ {0, 0.4, 0.6, 0.8}, where s = 0 is the balanced setting, and each setting is repeated 20 times. The conclusions hold under every allocation.

  1. Monotonicity in speaker count. Both metrics remain strongly monotonic in the number of speakers. For s = 0, 0.4, 0.6 and 0.8, Spearman ρ is 0.954, 0.951, 0.947 and 0.935 for SRS and −0.878, −0.869, −0.864 and −0.837 for FSD, with all p < 10⁻⁵⁰.
  2. Sensitivity to skew. At a fixed speaker count, a more uneven allocation lowers SRS and raises FSD, and mean SRS is ordered s = 0 > 0.4 > 0.6 > 0.8 at every speaker count.

A skewed allocation concentrates speech on a few dominant timbres and lowers effective diversity, so both metrics capture how evenly utterances are spread as well as how many speakers are present.

Sensitivity to the uniform-occupancy assumption. With 10,000 generated samples, the extrapolation changes the score very little. Averaged over 20 runs, the largest gap D̂ − k between the extrapolated support D̂ and the observed number of distinct retrieved entries k is 29.7 entries, which is 0.30% of the reference. This occurs at s = 0 with 1,960 speakers, where k = 2,276.5 and D̂ = 2,306.2. The reported SRS is therefore driven mainly by the observed retrieval coverage. The assumption also errs in the conservative direction. For a given support, uniform sampling yields the most distinct entries. A skewed generator repeats entries more often, which pushes k and D̂ downward rather than inflating them, so uneven generation is penalized, consistent with the lower SRS at larger s. The revision states that neither FSD nor SRS is a direct count of perceptually distinct speakers (Section 3.3.2, page 4, lines 205–207). It also states that the insertion experiment establishes monotonic sensitivity rather than a calibrated mapping to speaker counts (Section 3.3.3, page 5, line 269).

AuthorsOpenReview ↗

Rebuttal by Authors (4/4)

Q3: Distinct speakers in the reference and the effect of residual duplication

Clustering threshold. We count timbre groups with a conservative criterion calibrated on AISHELL-2, which provides speaker labels. With CAM++ embeddings, utterances of the same speaker have a mean cosine similarity of 0.710 (s.d. 0.101). The most similar cross-speaker utterance pair, measured over 200 random speakers, has a mean similarity of only 0.549 (s.d. 0.093). A threshold of 0.7 therefore groups only voices that are acoustically very close.

Group counts. Average-linkage clustering of the 10,000 reference embeddings at this threshold yields 9,555 timbre groups (Appendix A.3, Table 5, page 17, lines 865–888).

Number of groupsSize 1Size 2Size 3Size 4Size 5
9,5559,22424563185

Only 7.8% of the utterances share a group with another utterance, and no group has more than five members. Emilia provides no speaker labels, so we cannot determine whether a highly similar pair comes from two distinct speakers with close voices or from the same speaker appearing in different recordings. In either case, such pairs reflect the non-uniform density of the empirical timbre distribution, so we keep them in the main evaluation (Appendix A.3, page 17, lines 890–897; Section 3.2.1, page 3, lines 152–155).

Effect on the scores. To measure the effect of residual duplication, we built 20 density-flattened references. Each keeps one utterance per group and refills the set with filtered Emilia utterances, preserving the size and the gender and age composition, so that no pair exceeds the threshold. We then recomputed FSD and SRS with the generated embeddings fixed (Table 6, pages 17–18, lines 898–933).

  • FSD. FSD changes by less than 0.13% for every model.
  • Self-retrieval ceiling. The ceiling drops from 0.823 to 0.814 because leave-one-out retrieval becomes harder. This raises SRS by a nearly uniform 1.14–1.16 percentage points for every model.
  • Generated coverage. Coverage decreases by 0.23–1.46 points for all models except Qwen3-TTS-GRPO, which gains 0.19. Generated speech, like the reference itself, retrieves fewer distinct entries once retrieval becomes harder, which confirms that the leave-one-out ceiling is an appropriate normalizer.
  • Rankings. The model ranking under both FSD and SRS is unchanged.

Residual duplication therefore does not affect any comparative conclusion.


Q3: Alignment to the selected reference versus real-world speaker diversity

We agree with this distinction, and the revision adopts it. The paper does not claim coverage of real-world speaker diversity in general. The reference is constructed as an empirical reference distribution of the speech variation observed in Emilia (Section 3.2.1, page 3, lines 137–152). It contains in-the-wild recordings with at most one segment per source recording, is balanced between male and female speakers with an additional subset of child speech, and is filtered to reduce the dominance of flat, neutral speech (Appendix A.2, page 16, lines 811–863). A lower FSD or a higher SRS therefore means, strictly, closer alignment with or broader coverage of this constructed reference. The FSD and SRS gains of GRPO are described in the same terms (Appendix A.8, page 22, lines 1184–1186).

This alignment is nevertheless informative about diversity. In the insertion experiment above, the evaluated sets differ only in how many speakers they contain and how evenly utterances are allocated. Against the same fixed reference, sets with more speakers and a more even allocation consistently obtain a lower FSD and a higher SRS. The AISHELL-2 speech is read speech that was not used to build the reference, so these trends reflect the number and allocation of speakers rather than a match to the reference's speaking style. More diverse distributions therefore achieve closer similarity to the reference.


We hope these responses address the reviewer's concerns and would be glad to provide further clarification.

Reviewer_4DrFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at timbre diversity in text-guided voice design, specifically whether a TTS model can generate different speaker identities when sampled multiple times with the same prompt. The authors build a benchmark called VoxDICE using 10,000 filtered utterances from Emilia and three levels of prompts (APS, DSD, and RP, following InstructTTSEval). They measure diversity at two levels using CAM++ embeddings: within-prompt variation (MPCD and Vendi Score) and global distribution coverage against the reference set (FSD and SRS), alongside instruction following rate (IFR) judged by Gemini-3.1-Pro. Evaluating seven open-source models shows that models with high instruction following usually suffer from low timbre diversity, and vice versa. To fix this, they apply GRPO to Qwen3-TTS with a gated reward: the model only gets the group-level timbre diversity bonus if the sample passes both a gender check and a Qwen3-Omni instruction check (using APS-reformulated prompts). The trained model improves on both diversity metrics and average IFR without hurting audio quality or CER.

Critical strengths
  • The problem setup makes a lot of sense for voice design. In real applications like character voice casting, getting the exact same voice across 10 random seeds makes a model much less useful, even if that single voice follows the prompt well. Looking at both within-condition variation and global coverage is a good way to frame the evaluation.
  • The authors put solid effort into validating their metrics instead of just proposing them. The speaker-insertion experiment on AISHELL-2 (Fig. 2) and the human perceptual listening tests for both within-condition diversity (5,850 pairs) and LLM judge agreement (2,700 pairs) make the benchmark credible.
  • The reward formulation in Section 4 is simple and practical. Multiplying the diversity bonus by the gender and instruction gates prevents the model from cheating the diversity metric by generating weird or wrong-gender voices. Rewriting DSD/RP prompts into APS format for the Qwen3-Omni reward model is also a nice practical trick, and the ablations in Table 6 clearly show that each piece is needed.
Critical weaknesses
  • The same speaker encoder (CAM++) is used both for the diversity reward during GRPO training and for all four diversity metrics at test time. When you directly optimize the policy to push CAM++ embeddings apart, there is a risk of overfitting to the specific embedding space rather than actually increasing perceived speaker diversity. The paper validates CAM++ metrics on existing models in Section 5.4, but doesn't evaluate the RL model with a held-out speaker verification model (e.g., ECAPA-TDNN or WavLM-SV) or run a direct human A/B test comparing Qwen3-TTS before and after GRPO.
  • Even after GRPO training, the absolute diversity of Qwen3-TTS (VS 3.86, SRS 0.104) is still much lower than baselines like Ming-omni-tts-0.5B (VS 3.96, SRS 0.227, FSD 0.430) or VoxCPM2 (VS 5.15, SRS 0.241). Also, on the RP (role-play) split, which is arguably the closest to the character casting use case motivated in the intro, IFR actually drops slightly on VoxDICE (50.0% to 48.4%). Since the RL method is only tested on Qwen3-TTS, it is hard to tell if it could also take a high-diversity model (like VoxCPM2 or MOSS) and bring its instruction following up to par.
  • In Section 3.2.1, the reference set is filtered from Emilia by keeping only clips with arousal > 0.8 and dominance > 0.8. This seems way too strict and biases the "real speaker distribution" toward loud, energetic, and dominant speech, effectively throwing out calm, soft, or low-arousal voices. Using this skewed subset as the ground-truth target for FSD and SRS might penalize models that generate softer speaking styles.
  • In Tables 1 and 2, the diversity scores are computed over all generated samples, including the ones that fail the instruction. Because of this, a model with ~50% IFR (like MiMo-Audio) gets a huge boost in diversity scores simply because half of its outputs ignore the prompt constraints.
  • The collision threshold description in Sections 4.3 and 4.4 is confusing. Eq. 9 defines a collision as cosine distance dij<τd_{ij} < \tau, and Section 4.4 says τ=0.7\tau = 0.7. However, in Table 1, the mean pairwise cosine distance (MPCD) of Qwen3-TTS is only around 0.28 to 0.36. If τ=0.7\tau = 0.7 is really cosine distance, then almost every pair in a rollout group would collide and DiD_i would be 0 almost all the time. It looks like τ=0.7\tau = 0.7 might actually be cosine similarity (which corresponds to cosine distance < 0.3).
  • Some dataset details are unclear in the main text. The prompt templates in Appendix A.6 are all in English, but Table 4 evaluates Pinyin CER and Table 6 mentions "on Chinese prompts". The language distribution of VoxDICE and the private RL training set should be stated clearly in Sections 3 and 4.
Questions for authors
  1. If you evaluate Qwen3-TTS vs. Qwen3-TTS-GRPO using an independent speaker encoder (such as WavLM-SV or ECAPA-TDNN) not seen during RL training, do the diversity gains still hold to the same degree?
  2. Could you clarify whether the collision threshold τ=0.7\tau = 0.7 in Section 4.4 is applied to cosine distance or cosine similarity? Also, how sensitive is the training to τ\tau and the ratio between the base reward (b=3b=3) and the diversity bonus (Di∈[0,1]D_i \in [0, 1])?
  3. Why did you filter the Emilia reference corpus with arousal > 0.8 and dominance > 0.8? Doesn't this bias the reference distribution away from calm or soft voices when computing FSD and SRS?
  4. Have you tried computing within-condition diversity only on the subset of samples that pass the instruction-following check, or applying your GRPO pipeline to one of the higher-diversity baselines (e.g., Ming-omni-tts or VoxCPM2)?
3 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We thank the reviewer for the careful and constructive assessment. We respond to each concern below. All section and table references below refer to the revised manuscript.


1. Potential overfitting to the reward encoder

The diversity gains persist with speaker encoders not used during training, but their magnitudes are smaller. We therefore acknowledge an encoder-specific component of the improvement rather than claiming that the gains transfer to the same degree.

We re-evaluate the same Qwen3-TTS and Qwen3-TTS-GRPO audio using ECAPA-TDNN and ERes2NetV2. Both encoders show improvements in every diversity metric across all three instruction scales. The following table summarizes the relative improvement in V-MPCD, which is computed only over compliant outputs. Relative improvements are reported because absolute values are not directly comparable across embedding spaces.

Evaluation encoderUsed for the RL diversity rewardAPS gainDSD gainRP gain
CAM++Yes24.36%20.17%7.17%
ECAPA-TDNNNo14.92%11.07%3.09%
ERes2NetV2No19.96%14.54%4.58%

FSD and SRS also improve under both independent encoders, indicating that the gains extend beyond the pairwise separation directly targeted by the reward.

These results are consistent with partial overfitting to CAM++, but the improvements do not disappear outside its embedding space. Together with the direct human comparison described below, they support an increase in timbre diversity beyond optimization of a single encoder. The evaluation protocol and complete results are provided in Appendix A.8 and Table 13, with the main finding summarized in Section 5.3.


2. Diversity among compliant outputs, direct human comparison, and the RP IFR decrease

We agree that diversity computed over all outputs cannot establish whether a model provides diverse alternatives that satisfy the instruction. The revised manuscript therefore distinguishes diversity over all generated outputs from diversity within the compliant subset and evaluates the latter both automatically and through a direct human comparison.

Automatic evaluation of compliant diversity. We introduce Valid MPCD, denoted V-MPCD, which computes the mean pairwise cosine distance only between instruction-compliant utterances generated from the same instruction and synthesis text. A case with fewer than two compliant utterances receives zero. Non-compliant outputs therefore make no contribution, and cases without a valid pair remain in the evaluation rather than being excluded. V-MPCD thus primarily reflects timbre separation within the compliant subset, although the zero assigned to cases without a valid pair means that it is not fully independent of the compliance rate. GRPO increases V-MPCD on every instruction scale. The RP result is the most informative in this respect: the automated IFR decreases from 50.0% to 48.4%, and fewer compliant outputs would, if anything, create more zero-valued cases and lower V-MPCD. V-MPCD nevertheless rises from 0.251 to 0.269, so this gain reflects greater separation among compliant outputs rather than a larger number of valid cases. The definition and results appear in Section 3.3.1, Table 1, and Section 5.3.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Direct human comparison of usable voice alternatives. Appendix A.10 compares Qwen3-TTS before and after GRPO using blinded human annotations under matched instructions and synthesis texts. Listeners first discard utterances that fail the instruction and then group the remaining utterances by perceived speaker identity. The resulting number of effective voices counts distinct compliant speakers, so neither an instruction-violating voice nor a repeated compliant timbre increases this count. Each model generates five utterances for each of 100 prompts per scale, and each group is assessed independently by three annotators.

The table below places the automatic and human results alongside each other. Effective voices are the mean number of distinct compliant speakers per group of five utterances. Human compliance is the percentage of utterances judged to follow the instruction. The two evaluations use their respective protocols and sample sets.

ScaleModelV-MPCD ↑Effective voices ↑Human compliance
APSQwen3-TTS0.2753.10476.2%
APSQwen3-TTS-GRPO0.3423.47080.6%
DSDQwen3-TTS0.3473.37779.4%
DSDQwen3-TTS-GRPO0.4173.76984.4%
RPQwen3-TTS0.2512.06251.2%
RPQwen3-TTS-GRPO0.2692.27750.4%

Every increase in effective voices remains significant after Holm correction, including RP. RP therefore shows improved usable diversity, not improved instruction following. Its human-judged compliance rate also decreases slightly, while the number of distinct compliant voices increases. This directly supports more perceptually distinct choices that satisfy the instruction, without attributing the gain to failed outputs or claiming improvement in every aspect of performance. The human evaluation protocol, confidence intervals, and statistical tests are reported in Appendix A.10 and Table 15.


3. Reference filtering and distributional bias

The filtering is intended to construct a reference with broad expressive coverage, rather than estimate the natural frequency of speaking styles in the unfiltered corpus. We acknowledge that selecting utterances by arousal and dominance changes the reference distribution. The relevant question is whether this selection produces a narrowly restricted reference or broadens the variation available for evaluating voice design.

The added analysis shows that the unfiltered Chinese Emilia sample is heavily concentrated in neutral speech. After filtering, the neutral proportion decreases from approximately 90% to 39%, while neutral speech remains substantially represented. The retained utterances also exhibit greater semantic variation in annotated emotion, pitch, speaking rate, and volume. These findings support broader attribute coverage, although they do not establish an unbiased population distribution or equal coverage of every calm or soft speaking style.

This distinction matters because the reference serves both to derive instructions and to evaluate global coverage. A reference dominated by homogeneous neutral speech could favor models that repeatedly generate similar voices and provide limited coverage of expressive voice-design requests. Our construction instead emphasizes expressive variation while preserving substantial neutral coverage and sampling across gender and age categories. Accordingly, FSD and SRS should be interpreted as alignment with and coverage of this constructed reference, not as universal measures of fidelity to all real-world speaking-style frequencies. The rationale and supporting analyses are now presented in Section 3.2.1 and Appendix A.2, including Tables 3 and 4.


AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

4. Absolute diversity and generalization beyond Qwen3-TTS

We agree that Qwen3-TTS-GRPO does not surpass the higher-diversity baselines in absolute diversity. Our claim is that instruction-gated GRPO improves the balance between diversity and instruction adherence within a model, rather than making Qwen3-TTS the most diverse system.

To test the method in a contrasting regime, we apply the same reward formulation and training data to MiMo-Audio-7B-Instruct, the most diverse baseline but one with relatively weak instruction following. The following table summarizes the average results across the three instruction scales. VS measures diversity over all generated outputs, whereas V-MPCD measures separation within the compliant subset.

ModelVS ↑IFR ↑V-MPCD ↑
Qwen3-TTS3.31969.7%0.291
Qwen3-TTS-GRPO3.86072.8%0.343
MiMo-Audio7.00950.8%0.451
MiMo-Audio-GRPO6.55360.6%0.496

For MiMo-Audio, IFR and V-MPCD improve on every scale. Raw MPCD and VS decrease moderately, and global coverage weakens, so this experiment does not show simultaneous improvement in every metric.

The same objective improves compliant diversity from two contrasting starting points. For Qwen3-TTS, timbre repetition is the principal limitation, and the diversity bonus encourages distinct compliant candidates. For MiMo-Audio, already diverse rollouts make instruction adherence the more restrictive requirement, and gating directs optimization toward compliant outputs. This interpretation is consistent with increased V-MPCD in both models despite their different changes in raw diversity. For MiMo-Audio, however, IFR rises by nearly ten percentage points, so part of its V-MPCD gain may come from fewer cases without a valid pair rather than from greater separation alone. The additional experiment supports applicability beyond Qwen3-TTS and across these two contrasting starting points, without establishing universal generalization to all backbones. The results and discussion are provided in Appendix A.9 and Table 14, and summarized in Section 5.3.


5. Sensitivity to reward hyperparameters

Training is sensitive to both the collision threshold and the relative strength of the base reward. Appendix A.7 examines distance thresholds of 0.2 and 0.4 around the default of 0.3, and base rewards of 1 and 5 around the default of 3.

A lower collision threshold makes the diversity bonus easier to saturate and weakens the incentive for additional timbre separation. A higher threshold increases raw diversity but reduces instruction following. Likewise, reducing the base reward gives diversity greater influence relative to gate satisfaction, whereas increasing it weakens the relative incentive for timbre separation. Both changes reduce V-MPCD compared with the default configuration.

The default achieves the highest V-MPCD on all three scales among the tested settings, supporting its balance between compliance and diversity. These results characterize the observed sensitivity rather than establish insensitivity to hyperparameter choice. The complete comparisons and interpretation are included in Appendix A.7 and Table 12.


6. Collision-threshold correction and language clarification

The collision criterion is defined on cosine distance, with a collision occurring when the distance is below τ\tau. The revised manuscript specifies the distance threshold as τ=0.3\tau = 0.3, equivalent to a cosine-similarity cutoff of 0.7. This corrects the original implementation-details statement of 0.7 and makes the numerical value consistent with the distance-based definition. The clarification appears in Sections 4.3 and 4.4.

VoxDICE is a Chinese benchmark, and the private GRPO training corpus is also Chinese. The English annotation templates do not indicate an English evaluation set. We have made the benchmark language explicit in Section 3.2.1 and the training language explicit in Section 4.4.


We hope these responses address the reviewer's concerns and would be glad to provide further clarification.

Reviewer_yAcrOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces VoxDICE, a benchmark for evaluating voice-design models in terms of instruction following and timbre diversity. It evaluates seven existing systems and observes that good instruction following does not necessarily lead to diverse voice outputs. The authors further propose a GRPO-based approach that encourages diversity while maintaining instruction constraints. Experiments with Qwen3-TTS show improvements in both aspects.

Critical strengths

[+] Evaluating diversity at both the individual-prompt and overall-distribution levels is a useful idea.

[+] The authors provide a fairly comprehensive evaluation, including comparisons with existing models and human studies.

Critical weaknesses

[-] I’m concerned about how to interpret the SRS coverage scores. The occupancy model assumes uniform sampling from an unknown set of reachable reference identities, but the validation uses controlled speaker sampling. This leaves open whether the model holds under the uneven speaker distributions we would expect from a generative system. It is also unclear how the reference corpus was checked for unique speaker identities or how representative it is. These issues raise doubts about whether the reported global coverage scores measure what the paper claims, which matters given that this benchmark is a central contribution.

[-] GRPO improves on the original Qwen3-TTS checkpoint, though its diversity scores are still well below those of some evaluated systems. My main concern is that both the reward and the evaluation rely on speaker embeddings. The method could therefore learn to increase differences in embedding space without producing voices that listeners perceive as having distinct speaker timbres. The human evaluation helps, but I would like clearer evidence that the gains reflect perceptually distinct speaker identities while preserving the requested vocal characteristics. As it stands, the evidence does not fully support the claim that the method reliably produces more diverse voices that also follow the instructions.

Questions for authors

I would prioritize the following questions because their answers could materially affect my assessment.

  1. How do you ensure that the reference dataset adequately represents the diversity of real speakers?

  2. How can you verify that the embedding-based metrics capture perceptually meaningful voice differences rather than simply acoustic variation?

  3. Have you considered testing the proposed GRPO method on other models to demonstrate its broader applicability?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We thank the reviewer for the thoughtful comments. We respond to each concern below. All section, page and line numbers refer to the revised manuscript.


W2 & Q2: Perceptual validity of the diversity gains and of the embedding-based metrics

We conducted a new human study that evaluates the GRPO gains without relying on the speaker encoder. The study measures timbre diversity and instruction following jointly. For each instruction scale, we sampled 100 prompts, and Qwen3-TTS and Qwen3-TTS-GRPO each generated five utterances per prompt. Annotators first discarded the utterances that did not follow the instruction and then grouped the remaining utterances by speaker identity. We refer to the number of resulting groups as the number of effective voices. Under this measure, a distinct voice counts only if it is compliant, and compliant voices that share a timbre are counted once. Each group was labeled by three annotators who did not know which model produced it.

ScaleEffective voices, Qwen3-TTSEffective voices, GRPOHolm-corrected pCompliance, Qwen3-TTSCompliance, GRPO
APS3.103.47< 3e-576.2%80.6%
DSD3.383.77< 3e-579.4%84.4%
RP2.062.281e-351.2%50.4%

GRPO significantly increases the number of effective voices on all three scales. Non-compliant utterances are excluded before speakers are counted, so the improvement cannot be attributed to outputs that deviate from the instruction. Human-judged compliance is preserved, and inter-annotator agreement on compliance is substantial on all scales. The full protocol and results are given in Appendix A.10 and Table 15 (pages 23–24, lines 1231–1284), and the study is summarized in Section 5.3 (page 9, lines 452–455).

The embedding-based metrics are supported by evidence at both scales.

  • Global scale. The speaker-insertion experiment varies the number of real speakers while within-speaker acoustic variation remains present. Both SRS and FSD change monotonically with the speaker count (Section 3.3.3, page 5, lines 251–269).
  • Local scale. VS and MPCD correlate strongly with human judgments of speaker identity, with Pearson correlations of 0.78 and 0.80 (Section 5.4, pages 9–10, lines 468–505). Listeners were instructed to judge speaker identity only and to ignore other paralinguistic attributes and speech content (Figure 6, Appendix A.13.1, page 26).
  • Perceptibility. A new two-alternative forced-choice study determines how large a score difference listeners can perceive. On Qwen3-TTS, GRPO raises VS by 0.54, which exceeds the perceptual threshold in the diversity tiers where the model lies. The improvement is therefore audible (Section 5.5, page 10, lines 508–527).
  • Independence from the reward encoder. We re-evaluated the outputs with ECAPA-TDNN and ERes2NetV2, neither of which is used in training. GRPO improves every metric on every scale and retains a substantial part of the gain measured with CAM++, roughly two thirds or more with ERes2NetV2 and roughly half with ECAPA-TDNN. This indicates some overfitting to CAM++, but the improvement persists in independent embedding spaces (Appendix A.8, Table 13, page 22, lines 1150–1187).

W1: Validity of SRS under uneven speaker distributions

The occupancy model serves only to extrapolate the observed coverage to the coverage attainable with unlimited generation (Appendix A.1, pages 14–15, lines 739–767). Whether SRS remains valid when speakers are used unevenly is therefore an empirical question, which we address directly. We expect a generator that concentrates its outputs on a few timbres to offer less usable diversity, and a valid metric should assign it a lower score.

To verify this empirically, we extended the speaker-insertion experiment from balanced allocations to Zipf allocations of increasing skew. Under every allocation, SRS increases monotonically with the number of speakers, with Spearman correlations between 0.935 and 0.954. At a fixed number of speakers, SRS decreases consistently as the allocation becomes more skewed. FSD shows the same pattern, although it does not rely on the occupancy model. SRS therefore reflects both the number of speakers and how evenly they are used. This experiment replaces the balanced-only validation in Section 3.3.3 and Figure 2 (page 5, lines 226–229 and 251–269). The revised text also states that SRS is a monotonic, distribution-level proxy for global diversity rather than a count of perceptually distinct speakers (Section 3.3.2, page 4, lines 204–206; Section 3.3.3, page 5, lines 268–269).


AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

W1 & Q1: Speaker uniqueness and representativeness of the reference corpus

Emilia does not provide speaker labels, so we sample at most one segment from each source recording. To detect any remaining near-duplicates, we cluster the reference embeddings with a conservative similarity threshold of 0.7. The threshold was calibrated on AISHELL-2, where it lies well above the similarity of the closest pairs from different speakers. The 10,000 reference utterances form 9,555 groups, and only 7.8% of the utterances share a group with another utterance. Without speaker labels, we cannot determine whether such a pair comes from two distinct speakers with similar voices or from the same speaker appearing in different recordings. In either case, these pairs reflect the non-uniform density of the empirical timbre distribution, so we keep them in the main evaluation. To assess their influence, we constructed 20 references that retain a single utterance per group while preserving the gender and age composition. On these references, FSD changes by less than 0.13% and SRS by at most 1.4% for every model, and the model ranking is unchanged. This analysis is presented in Appendix A.3, Tables 5–6 (pages 17–18, lines 865–933), and referenced in Section 3.2.1 (page 3, lines 152–155).

The reference is intended as a common real-speech yardstick for Mandarin voice design. It is drawn from in-the-wild Emilia recordings, covers 9,555 timbre groups, and is balanced between male and female speakers, with an additional subset of child speech. Our filtering broadens the distribution rather than narrowing it. Neutral speech accounts for nearly 90% of raw Emilia but 39% of the reference, and angry, happy, surprised and sad speech are all represented. The resulting voice-attribute descriptions are also substantially more diverse. In addition, SRS continues to increase as speakers from AISHELL-2 are added, although this corpus was not used to build the reference. The reference is therefore broad enough to distinguish different levels of diversity. The construction is described in Section 3.2.1 (page 3, lines 137–153), and the emotion and attribute analyses are given in Appendix A.2, Tables 3–4 (page 16, lines 811–863). We also note that the reference represents Mandarin speech only.


Q3: Applicability of GRPO to other models

We applied the same GRPO procedure to MiMo-Audio-7B-Instruct, keeping the reward and training data unchanged. This model represents the opposite case to Qwen3-TTS: it is the most diverse baseline but one of the least compliant. After training, average IFR increases from 50.8% to 60.6%, and V-MPCD, which measures diversity among compliant outputs, increases from 0.451 to 0.496. Both metrics improve on all three scales. Raw diversity decreases moderately, yet MiMo-Audio-GRPO still achieves the best average MPCD, VS, FSD and SRS among all evaluated models.

This behavior follows from the instruction gate. A distinct but non-compliant voice receives a lower reward than a compliant one, so for a model that is already highly diverse, the gate dominates the policy update. For Qwen3-TTS, whose main limitation is timbre repetition, the same objective primarily increases diversity. In both cases, compliant diversity improves, which indicates that the method addresses whichever capability limits the base model. The full results are reported in Appendix A.9 and Table 14 (page 23, lines 1189–1230) and summarized in Section 5.3 (page 9, lines 463–465).


We hope these responses address the reviewer's concerns and would be glad to provide further clarification.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks to the insightful and professional reviewers.

Dear all reviewers,

We sincerely thank all reviewers for their constructive comments and the time they devoted to our work. We are greatly encouraged that the reviewers found the problem setting well motivated for voice design and recognized the effort we put into validating the metrics.

In response to the reviewers' questions and suggestions, we have conducted new experiments and revised the paper. The main additions are summarized below, and detailed answers are given in the individual responses.

Perceptual validity of the GRPO gains. To check that the gains do not merely reflect overfitting to the reward encoder, we re-evaluated the outputs with ECAPA-TDNN and ERes2NetV2, neither of which is used in training. Every diversity metric still improves on every scale, although by a smaller margin than under CAM++. We also ran a blinded human study in which annotators discard non-compliant utterances and then count distinct speakers among the remaining ones. GRPO significantly increases the number of these effective voices on all three scales (e.g., 3.10 → 3.47 on APS, Holm-corrected p < 3×10⁻⁵), while human-judged compliance is preserved. A new 1500-pair 2AFC study further shows that the VS and MPCD gains of GRPO exceed the perceptual thresholds of the tiers in which Qwen3-TTS lies (Section 5.5, Appendices A.8, A.10 and A.12).

Diversity among instruction-compliant outputs. We introduce Valid MPCD (V-MPCD), which measures diversity only among compliant outputs under a fixed generation budget. GRPO raises V-MPCD on every scale, including RP (0.251 → 0.269), where IFR decreases slightly. After the judge's errors are propagated through a two-stage bootstrap, the 95% intervals of the two models remain non-overlapping on all scales (Section 3.3.1, Table 1, Appendix A.5).

Validity of the global metrics and the reference set. We extended the speaker-insertion experiment to Zipf-skewed speaker allocations. SRS and FSD remain strongly monotonic in the number of speakers (|ρ| ≥ 0.935 for SRS), and at a fixed speaker count both metrics penalize more uneven allocations. A duplication audit groups the 10,000 reference utterances into 9,555 timbre groups, and on 20 de-duplicated references FSD changes by less than 0.13%, SRS by at most 1.4%, and the model ranking is unchanged. We also show that the arousal and dominance filter broadens rather than narrows the reference, reducing neutral speech from about 90% to 39% (Section 3.3.3, Appendices A.2–A.3).

Applicability of GRPO beyond Qwen3-TTS. We applied the same reward and training data to MiMo-Audio-7B-Instruct, the most diverse but one of the least compliant baselines. IFR increases from 50.8% to 60.6% and V-MPCD from 0.451 to 0.496, so compliant diversity improves from two contrasting starting points (Appendix A.9, Table 14).

Reproducibility and clarifications. We released an anonymous evaluation package (https://huggingface.co/datasets/ICLR45502/VoxDice_anonymous) and reported the annotation protocol, inter-annotator agreement and the confusion matrix of the IFR judge. We also added a sensitivity analysis of the reward hyperparameters, corrected the collision threshold, and clarified that VoxDICE and the training corpus are in Chinese.

We have revised the paper according to the reviewers' suggestions. We believe that these additional experiments and revisions strongly support the effectiveness and contribution of both the VoxDICE benchmark and the proposed RL method. We welcome any further feedback from the reviewers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.