{"id":"4282c0e3-2d0c-4066-bcae-d345eb50c1f3","arxiv_id":"2505.04113","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Preference alignment on the new INTP dataset improves intelligibility and quality of zero-shot TTS across diverse domains, with weak-to-strong generalization shown on CosyVoice 2 and Ints.","lead":"The paper introduces INTP, a 250K-pair preference dataset for zero-shot text-to-speech, and applies Direct Preference Optimization to improve intelligibility in tongue twisters, repeated words, code-switching, and cross-lingual synthesis. It reports consistent intelligibility and quality gains across five TTS models, including two stronger models that saw no INTP data during construction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Code-switching WER is measured by Paraformer-zh, a Mandarin-trained ASR, in both INTP's preference labels and Table 4; human verification (Table 3) covers only regular texts, so the large code-switching WER gains may reflect ASR adaptation rather than real intelligibility improvement.","rationale":"The paper's central claim requires that DPO with INTP improves human-level intelligibility in articulatory, code-switching, and cross-lingual domains, with the headline evidence being the WER deltas in Table 4. For that to hold, the WER instrument must be a valid measure of intelligibility in each domain and must not be gameable by the same optimization procedure that uses it as the label source. Section 3.3 constructs preference pairs by ranking samples with WER (intra and inter) and Section 5 evaluates with the same two ASR models; this labeler-judge identity is the structural weakness. The sharpest failure is in code-switching: Paraformer-zh, a Mandarin-trained recognizer, is used on English-Chinese mixed text. A Mandarin ASR cannot reliably transcribe the English tokens, so (i) base WERs of 30-54% in this domain overstate TTS error, and (ii) DPO can lower measured WER by shifting the model's English output toward what this specific ASR prefers (e.g., Chinese-accented or reduced-salience English), which is not the same as greater human intelligibility. The reported large gains (F5-TTS 33.99 to 15.98) are exactly where this mechanism bites. The paper's human checks do not rescue the code-switching sub-claim: Table 3's 300-pair verification is on regular texts only, and Figure 2a's reading-accuracy sample pools domains at roughly 80 samples per model, so code-switching-specific human confirmation is absent. The fix is inexpensive: recompute the domain's WER with a bilingual ASR (or a code-switched ASR benchmark) and, if feasible, a per-domain human reading-accuracy test. If the deltas persist under the bilingual ASR and agree with human judgments, the code-switching claim stands; if not, the paper should narrow the claim to domains whose instruments are valid. Other weaknesses (no seed variance or significance testing; abstract 'similarity' claim vs F5-TTS SIM drops in Table 4) reinforce the conditional posture but are secondary. I therefore keep the reader's CONDITIONAL verdict (UNCHANGED) and agree with the reader's identification of the WER-as-proxy assumption, sharpened to a domain-specific instrument-validity failure.","tokens_in":28591,"tokens_out":17376,"duration_ms":163099,"concrete_test":"Recompute the code-switching rows of Table 4 for all five models using Whisper-large-v3 (or another bilingual ASR) on the full 1,000-sample en-zh code-switching evaluation set, comparing base versus aligned WER deltas. If the deltas shrink to near zero or reverse under the bilingual ASR, the code-switching component of the central claim is an artifact of Paraformer-zh's inability to score English segments. In the same protocol, add a domain-stratified human reading-accuracy test on at least 100 code-switching samples per model (base vs aligned) and measure the correlation between Paraformer-zh WER and human reading accuracy on those samples; a weak or negative correlation would show that this instrument is invalid for the code-switching domain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Both the INTP preference labels and the Table 4 evaluation rely on the same two ASR instruments (Section 5: Whisper-large-v3 for English, Paraformer-zh for Chinese and code-switching), and for one of the paper's three headline domains this instrument cannot measure the quantity it is supposed to score. Paraformer-zh is a Mandarin-trained recognizer; on English-Chinese mixed text such as '熊猫吃shoots和leaves', the English tokens are systematically misrecognized or deleted, so the reported WER conflates TTS error with fixed ASR-instrument error. Consequently the very high base code-switching WERs (ARS 54.15%, F5-TTS 33.99%, Table 4) are partly an artifact, and, more importantly, the DPO alignment can reduce these numbers by making English segments more 'Paraformer-friendly' (e.g., Chinese-like articulation) without improving human intelligibility. The claimed gains in this domain (F5-TTS 33.99 to 15.98; ARS 54.15 to 36.91; CosyVoice 2 33.36 to 28.31) are thus not established as genuine intelligibility improvements. The paper's human verification does not cover this: the 300-pair reading-accuracy check in Section 3.4 is explicitly 'on INTP regular texts' only, and the Figure 2a reading-accuracy sample (~80 samples per model) is small and pooled across domains, so it cannot confirm or refute the code-switching WER numbers. The regular, articulatory, and cross-lingual rows use language-appropriate ASRs and are less exposed, but the identical ASR-as-labeler-and-judge design leaves all headline deltas vulnerable in principle. A secondary internal inconsistency (abstract claims 'similarity' improvements while Table 4 shows F5-TTS SIM dropping from 0.615 to 0.591) supports, but is not, the central concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes preference alignment to improve zero-shot text-to-speech (TTS) intelligibility across regular, articulatory, code-switching, and cross-lingual domains. It introduces INTP, a synthetic preference dataset of roughly 250K pairs built from ARS, F5-TTS, and MaskGCT, with preference pairs constructed via intra-model sampling, inter-model comparisons, and human-guided prompt perturbations, all labeled using WER. The authors extend DPO to autoregressive, flow-matching, and masked-generative TTS architectures and report WER, SIM, UTMOS, N-CMOS, and subjective reading-accuracy improvements for the three INTP models and for two stronger models (CosyVoice 2 and Ints) not used in INTP construction, including an iterative alignment round for Ints. The central empirical claims are summarized in Table 4 and Figure 2.","tokens_in":1513,"tokens_out":1774,"duration_ms":82633,"significance":"If the results hold, the contribution is significant: the paper provides a large public preference dataset for TTS intelligibility, demonstrates DPO variants for non-autoregressive TTS architectures, and presents evidence of weak-to-strong generalization and iterative alignment. The release of the dataset, code, models, and evaluation sets, together with machine-checkable preference data, is a clear strength. The human reading-accuracy check on regular texts (Table 3) supports the validity of WER-based preferences in that domain, and the consistent objective and subjective trends for regular, articulatory, and cross-lingual cases are encouraging. The main unverified risk is the code-switching WER instrument issue, which directly affects one of the paper's headline domains.","major_comments":[{"comment":"The code-switching WER numbers are produced by Paraformer-zh, a Mandarin-centric ASR, on English-Chinese mixed text, and the same instrument is used both to label INTP preferences and to evaluate the aligned models. On mixed-language text, this ASR systematically misrecognizes or deletes English tokens, so the high baseline WERs (e.g., ARS 54.15, F5-TTS 33.99 in Table 4) include a large fixed instrument-error component, and alignment can reduce WER by making English segments more Paraformer-friendly without improving human intelligibility. The claimed code-switching gains (e.g., F5-TTS 33.99 to 15.98, ARS 54.15 to 36.91, CosyVoice 2 33.36 to 28.31) are therefore not established as genuine intelligibility improvements. Because the human verification in Section 3.4 is explicitly on INTP regular texts only, this domain currently has no human ground truth. Please add a human reading-accuracy evaluation on a code-switching subset, or re-measure with an ASR that handles mixed-language input, and report the preference-labeling and evaluation ASRs separately.","section":"Section 5, Table 4, Appendix F.2"},{"comment":"WER-based preference labels are human-verified only for regular texts (Table 3), while the same ASR models used to rank positive and negative samples in INTP are also used to score the final models in Table 4 for every domain. For articulatory, code-switching, and cross-lingual data, there is no per-domain human verification and no analysis of how ASR label noise propagates into the DPO training signal. This is load-bearing for the paper's diverse-domains claim because it leaves open the possibility that alignment optimizes ASR-specific artifacts rather than intelligibility in these non-regular domains. A human verification on a few hundred pairs in each non-regular domain, even if smaller than the regular-text verification, would directly address this risk.","section":"Sections 3.3-3.4"},{"comment":"The subjective evaluation reporting is internally inconsistent: Section 5.1 states that we randomly sample 300 samples for subjective evaluation, while Appendix F.3 describes 10 samples per language per setting, yielding 80 pairs per model and 400 pairs across the five systems. This mismatch prevents the reader from knowing how many listening judgments actually underlie Figure 2. With roughly 80 samples per model pooled across four domains, the reading-accuracy results in Figure 2a are also too coarse to confirm domain-specific claims, especially the code-switching claim. Please reconcile the sample counts and, if possible, report per-domain reading-accuracy values with uncertainty estimates.","section":"Section 5.1 and Appendix F.3"}],"minor_comments":[{"comment":"The N-CMOS column shows values only for the aligned models and '-' for the base models, which is ambiguous; please state explicitly that N-CMOS is a paired comparison of the aligned model against its base model and describe the sign convention.","section":"Table 4 caption"},{"comment":"The text says the scripts for repeated-text and punctuation-perturbation construction will be included in the future open-source repository, which conflicts with Section 1's claim that all resources are already open-sourced; if the scripts are available, please provide the link.","section":"Appendix A.1"},{"comment":"The TTS Intelligibility Arena table is hard to read because the Win Rate column sums row percentages but the diagonal entries are omitted; please add a short description of how the win rates are computed from the 10K inter-pairs and what the percentages in each cell mean (row versus column win rates).","section":"Table 2"},{"comment":"Figure 2a reports reading accuracy without error bars or confidence intervals; given the small sample size, please add uncertainty estimates or state the number of judgments per bar.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations and the non-regular-domain human verification gap is real. The code-switching WER issue is the main correctness risk and should be resolved with human or alternative-ASR evidence before publication. The regular-domain and cross-lingual results are considerably more secure, so I see this as fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nBottom line: this is a genuinely useful resource—a 250K-pair preference dataset for TTS intelligibility, open-sourced, plus DPO variants for flow-matching and masked generative models—and the experiments mostly support the headline. I’d send it out, but I’d make the authors qualify one big claim before acceptance.\n\nWhat’s new: INTP is the first intelligibility preference set that mixes intra-model, inter-model, and human-guided perturbed pairs at this scale. The DPO extensions are straightforward applications of Wallace et al. to OT-FM and MGM, and they work: WER drops consistently across ARS, F5-TTS, MaskGCT, CosyVoice 2, and Ints in regular, articulatory, and cross-lingual domains. The weak-to-strong generalization on CosyVoice 2 and Ints is a nice result, and the iterative alignment round (Ints v2) shows the flywheel concept. I also credit them for releasing data and code.\n\nThe soft spot is code-switching. The paper uses Paraformer-zh, a Mandarin-only ASR, to both build the preference pairs and evaluate the final models for mixed English-Chinese text. That’s a circular instrument for this domain: the reported WER conflates TTS errors with ASR’s fixed inability to transcribe English tokens. So the large gains (F5-TTS from 34 to 16, ARS from 54 to 37, CosyVoice 2 from 33 to 28) are not yet established as real intelligibility gains. They could partly be the models learning to produce English that Paraformer finds easier to swallow. The human verification is done on regular texts only, so it doesn’t save this domain. That problem doesn’t sink the other three domains—they use language-appropriate ASRs—but it needs a bilingual ASR (or human listening) on the code-switching set before I’d trust those numbers.\n\nTwo smaller issues. The abstract says ‘similarity improvement,’ but Table 4 shows SIM drops for F5-TTS across the board, and the A/B test is mostly ties. That claim should be toned down. And there are no seed variances or significance tests anywhere; given the moderate effect sizes, that’s a real gap.\n\nOverall, the core intelligibility and naturalness results for regular, articulatory, and cross-lingual cases look solid enough to build on. The paper deserves a serious peer review, and I’d want the revision to fix the code-switching evaluation and soften the similarity claim.","headline":"Useful dataset and DPO recipe for TTS intelligibility, but the code-switching win is probably inflated by a Mandarin-only ASR scoring both the labels and final results.","tokens_in":29546,"tokens_out":3023,"would_cite":true,"duration_ms":29405,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fine-tuning zero-shot TTS on a 250K-pair preference dataset cuts word errors on tongue twisters, code-switching, and cross-lingual speech, and transfers to stronger models that never saw the data.","keywords":["zero-shot text-to-speech","intelligibility","preference alignment","direct preference optimization","speech preference dataset","code-switching","cross-lingual synthesis","weak-to-strong generalization"],"falsifier":"Take a stratified sample of the roughly 251K INTP pairs — especially code-switching and cross-lingual ones — and have native listeners (or a second, independent recognizer) transcribe both members of each pair; if a sizable share of pairs reverse their word-error ranking, the preference labels are partly recognizer noise and the intelligibility claim is weakened. A second check: rebuild the preference pairs using only CosyVoice 2's own outputs and realign it; if the gains vanish, the transfer depends on cross-model diversity rather than on the preference signal itself.","tokens_in":28330,"feed_emoji":"🗣️","tokens_out":10246,"duration_ms":86002,"temperature":0.7,"pith_summary":"Modern zero-shot text-to-speech systems read ordinary sentences fluently but frequently garble tongue twisters, repeated words, code-switched text, and cross-lingual speech. This paper's claim is that those failures can be treated as a preference-learning problem: build a large dataset whose \"good\" and \"bad\" speech clips are ranked by word error rate plus deliberately crafted human-guided errors, then fine-tune a TTS model with Direct Preference Optimization adapted to its architecture. The authors report that after this alignment, five different TTS models — including two stronger models that never contributed data to the dataset — improve on intelligibility, naturalness, and overall audio quality across all four difficult domains. They also report that a second round of alignment on data generated by the aligned model yields further, smaller gains, and that the preference dataset doubles as a tool for ranking competing TTS models. If correct, the approach gives a cheap, data-only recipe for fixing failure modes that current large-scale pre-training leaves unresolved.","feed_headline":"Preference alignment slashes TTS errors on tricky speech","feed_subtitle":"A 250,000-pair dataset fixes garbled tongue twisters, code-switching, and cross-lingual speech in five TTS models.","key_machinery":"The load-bearing object is INTP, a preference dataset of roughly 251K pairs constructed by three complementary routes: intra-pairs (several samples of the same model on the same prompt, ranked by word error rate), inter-pairs (the same prompt synthesized by different models, so their complementary strengths can be pooled and, as a side effect, an \"intelligibility arena\" that ranks models), and perturbed pairs, in which the target text itself is deliberately corrupted — misspelled into near-homophones or re-punctuated to shift meaning — to create human-guided negative samples. The optimization machinery is Direct Preference Optimization (DPO), which converts preference pairs into a policy-gradient-free loss using the closed form of the RLHF objective; the paper derives two new variants, DPO-FM for flow-matching models, expressed as a velocity-field difference on optimal-transport interpolation paths, and DPO-MGM for masked generative models, expressed over the log-probabilities of unmasked tokens. Word error rate from Whisper-large-v3 (English) and Paraformer-zh (Chinese and code-switching) is the ranking signal that turns raw audio into preferences.","core_discovery":"The paper's central claim is that the intelligibility failures of zero-shot TTS in hard domains are an out-of-distribution problem that preference alignment can repair without new pre-training data. Concretely, it claims that Direct Preference Optimization fine-tuned on INTP — about 251,000 triples of (prompt, preferred speech, dispreferred speech) spanning regular, repeated, code-switching, pronunciation-perturbed, and punctuation-perturbed Chinese and English texts — lowers word error rate while also improving naturalness (N-CMOS) and overall quality (UTMOS) for autoregressive, flow-matching, and masked-generative models, both for the three models that generated the data (ARS, F5-TTS, MaskGCT) and for two stronger held-out models (CosyVoice 2, Ints). The improvements on the held-out models are presented as weak-to-strong generalization: preference data harvested from weaker models transfers to stronger ones. A final claim is that iterative alignment — letting the aligned Ints model generate a fresh round of preference pairs — produces further gains, with a smaller second-round effect that the authors attribute to the base model's inherent ceiling.","pith_inferences":["The paper leaves implicit that the same pipeline applies to other failure modes — rare proper nouns, technical jargon, low-resource language pairs — but the WER-based ranking signal would need re-validation there, since a recognizer that cannot spell a word will mislabel both sides of a pair.","A consequence of the design worth testing: because WER ranks both members of every pair, the ceiling on dataset quality is set by the recognizers' own blind spots; an audit that re-transcribes a stratified sample with human listeners would quantify how much of the measured gain is preference signal versus recognizer bias.","The inter-pair arena could double as a continuously updated public benchmark for TTS intelligibility rather than a one-off training set, since its model rankings already align with human perception.","Reading the diminishing second-round gains as a capacity ceiling suggests the next bottleneck is the base model itself; combining INTP-style alignment with pre-training that actually includes the hard domains would likely outperform either alone — a combination this paper does not test."],"forward_implications":["Any zero-shot TTS model — autoregressive, flow-matching, or masked-generative — can be made more intelligible on hard domains by fine-tuning on a preference dataset, with no change to its architecture.","Preference datasets do not need to be built with the strongest available model; data from weaker models transfers to stronger ones, lowering the cost of dataset construction.","Iterative alignment provides a data-and-model flywheel: the aligned model generates the next round of preference pairs, yielding further gains until the base model's own capability becomes the bottleneck.","Failures are domain-specific — repeated-word data helps articulatory cases, pronunciation-perturbed data helps cross-lingual cases — so post-training data can be targeted instead of generic, and combining all data types gives the best overall result.","Because alignment also raised naturalness and kept speaker similarity roughly constant, intelligibility preference data does not appear to trade off against other quality dimensions.","The inter-pair construction also yields an intelligibility ranking of models that agrees with human reading-accuracy judgments, so the dataset doubles as an evaluation tool."],"supporting_citations":[{"why":"Supplies the Direct Preference Optimization loss; every alignment variant in the paper (vanilla, flow-matching, masked-generative) adapts this closed-form reward reparameterization.","marker":"[Rafailov et al., 2023]"},{"why":"Defines ARS and MaskGCT, two of the three models used to synthesize INTP, and provides the masked-generative architecture the MGM-DPO variant targets.","marker":"[Wang et al., 2025a]"},{"why":"F5-TTS, the flow-matching model used both to synthesize INTP data and as the testbed for the proposed DPO-FM extension.","marker":"[Chen et al., 2024c]"},{"why":"CosyVoice 2, the held-out stronger model that tests INTP's weak-to-strong generalization and is the best-performing comparison in the inter-pair intelligibility arena.","marker":"[Du et al., 2024b]"},{"why":"Emilia(-Large), the corpus whose speech and transcriptions are stratified-sampled to build INTP prompts.","marker":"[He et al., 2024]"},{"why":"Whisper-large-v3, the recognizer whose word error rate both ranks preference pairs and measures final intelligibility on English.","marker":"[Radford et al., 2023]"},{"why":"Paraformer-zh, the recognizer whose word error rate ranks Chinese and code-switching pairs and evaluations.","marker":"[Gao et al., 2022]"},{"why":"DeepSeek-V3, the language model that cleans source texts and fabricates pronunciation- and punctuation-perturbed negative texts.","marker":"[DeepSeek-AI et al., 2024]"},{"why":"The diffusion-DPO derivation that the paper adapts to obtain its flow-matching DPO loss.","marker":"[Wallace et al., 2024]"},{"why":"The prior WER-based intra-pair preference pipeline that INTP extends with inter-model and perturbed pairs.","marker":"[Tian et al., 2024]"}],"fun_headline_variants":["250k preference pairs fix TTS tongue twisters","Weak-to-strong TTS boost via preference alignment","DPO on 251k pairs makes TTS clear on hard speech","Preference alignment makes TTS intelligible in tough domains","Iterative alignment further improves TTS intelligibility"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on assuming that two automatic speech recognizers' word-error counts correctly say which of two clips is more intelligible — an assumption checked by human listeners on only 300 of about 251,000 pairs, so if the recognizers mis-hear accented or code-switched speech in a patterned way, the dataset's preferred and dispreferred labels silently inherit that error.","fun_headline_variants_meta":{"raw":{"variants":["250k preference pairs fix TTS tongue twisters","Weak-to-strong TTS boost via preference alignment","DPO on 251k pairs makes TTS clear on hard speech","Preference alignment makes TTS intelligible in tough domains","Iterative alignment further improves TTS intelligibility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000996,"raw_usage":{"total_tokens":4231,"prompt_tokens":974,"completion_tokens":3257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":3191}},"tokens_in":590,"tokens_out":3257,"duration_ms":22059,"temperature":1.0,"reasoning_tokens":3191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:35:44.128438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a stratified sample of the roughly 251K INTP pairs — especially code-switching and cross-lingual ones — and have native listeners (or a second, independent recognizer) transcribe both members of each pair; if a sizable share of pairs reverse their word-error ranking, the preference labels are partly recognizer noise and the intelligibility claim is weakened. A second check: rebuild the preference pairs using only CosyVoice 2's own outputs and realign it; if the gains vanish, the transfer depends on cross-model diversity rather than on the preference signal itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Paraformer-zh, the recognizer whose word error rate ranks Chinese and code-switching pairs and evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The diffusion-DPO derivation that the paper adapts to obtain its flow-matching DPO loss."}],"review_version":1}