Pith. sign in

REVIEW 4 major objections 4 minor 115 references

Experience-Calibrated Contrastive Decoding for Mitigating Hallucinations in LM-Based Text-to-Speech

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A training-free decoding method that contrasts a speech LM's text-conditioned and text-free next-token predictions cuts content errors in LM-based TTS by up to 55.6% across four models and in 24 of 25 multilingual settings.

desk verdict A training-free decoding-time fix for LM-TTS hallucinations that mostly works and deserves real referees once the missing error bars are addressed. read the letter →

arxiv 2608.00722 v1 pith:LWBJZFRI submitted 2026-08-01 eess.AS cs.LGeess.SP

Chenlin Liu , Minghui Fang , Zhonghao Bi , Zekai Su , Rong Wang , Jiqing Han This is my paper · ORCID
classification eess.AScs.LGeess.SP
keywords speechhallucinationcontrastivedecodingLM-basedtext-to-speechdecoding-timecontrolalignmentinformationexperienceacoustictokengenerationtraining-freemitigation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a substantial class of speech hallucinations in language-model text-to-speech begins as a local shortfall in text-derived alignment support at a vulnerable token transition, and that this shortfall can be corrected at decoding time without retraining. It introduces Experience-Calibrated Contrastive Decoding (ECCD), which contrasts the same speech model's next-token predictions with and without text conditions, keeps the original expert distribution as an anchor, and adds only positive alignment enhancement, scaled by an experience-compatibility coefficient. Across four speech LMs, ECCD lowers word/character error rates by up to 55.6% in all SeedTTS-Eval settings and in 24 of 25 multilingual CV3-Eval settings, with a listening-test CMOS gain of +0.644. If these results hold, decoding-time conditional-information control is a viable complement to architecture- and training-based hallucination mitigation.

What carries the argument

The load-bearing object is the ECCD score of Eq. (8): $S_{\text{ECCD}}(x_i)=\log p_E(x_i)+\alpha(1-C_i)[\log(p_E(x_i)/p_A(x_i))]_+$ inside the expert-defined top-$k$ plausible set $V_{\text{head}}$, and simply $\log p_E(x_i)$ outside it. The contrast between the full-conditioned expert $p_E$ and the text-ablated amateur $p_A$ — same model, same acoustic history, only the text conditions removed — is the operational signal of incremental alignment support. $C_i$, the experience compatibility coefficient, is the total probability mass $p_A$ assigns to the top-$k$ expert-plausible set: a large $C_i$ means the experience proxy already agrees with the expert's candidates, so the intervention is attenuated, while a small $C_i$ triggers stronger positive enhancement. The design deliberately avoids conventional contrastive decoding's move of treating amateur support as negative evidence, because $p_A$ carries pronunciation, prosody, duration, and continuity information that must be preserved for fluent speech.

What would settle it

Compute $I_i$ and $G_i$ on a large set of utterances with reliable ASR labels and compare first-error boundaries with matched correct boundaries under the same token-rate mapping; if the matched correct-minus-error differences in $G_i$ do not replicate or flip sign, the onset-shortfall claim fails. Alternatively, apply ECCD to a corpus where text removal is known to alter prosody without causing content errors: if the CER gains disappear or reverse while lower-temperature sampling still helps, the alignment-support story is wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that hallucination onset in LM-based TTS occurs when the selected acoustic token receives insufficient relative alignment support from the text, even when distribution-level text influence is present; once the off-target token enters the history, the error propagates. Operationally, the paper defines the full-condition distribution $p_E(x_i)=p(x_i\mid x_{<i},a,T_p,T_t)$ and the text-ablated distribution $p_A(x_i)=p(x_i\mid x_{<i},a)$, and treats their discrepancy as alignment influence. ECCD then modifies the next-token score inside an expert-defined top-$k$ set: $S_{\text{ECCD}}(x_i)=\log p_E(x_i)+\alpha(1-C_i)[\log(p_E(x_i)/p_A(x_i))]_+$, with the plain expert score $\log p_E(x_i)$ used outside the set. The method keeps the expert distribution as an anchor, applies only the positive part of the contrast, and scales that correction by $(1-C_i)$, where $C_i$ is the total probability mass $p_A$ assigns to the expert-plausible set. The paper reports that this reduces content errors across four models in all SeedTTS-Eval subsets and in 24 of 25 CV3-Eval language/subset combinations, and that the per-token measures $I_i$ (forward KL from expert to ablated distribution) and $G_i$ (the selected token's log-ratio) are lower at first-error boundaries than at matched correct boundaries, supporting an onset-specific alignment shortfall.

Load-bearing premise

The load-bearing premise is that the text-ablated prediction $p_A(x_i \mid x_{<i}, a)$ is a faithful operational proxy for experience-oriented prediction, so that the contrast $\log(p_E/p_A)$ isolates text-derived alignment support; if dropping the text instead changes prosody, silence, or generative quality in ways unrelated to alignment, the enhancement and the measured $I_i$ and $G_i$ signatures could be chasing an artifact.

Editorial extensions

If this is right

  • ECCD can be applied to any autoregressive acoustic-token language model without architectural changes or additional training, since it only modifies next-token scores before the model's native filtering and penalty operations.
  • The method transfers across languages and speaker conditions: error reductions appear in 24 of 25 model/subset combinations across nine languages, suggesting that decoding-time alignment enhancement generalizes beyond a single training distribution.
  • Lower-temperature sampling does not reproduce ECCD's gains and even raises test-en word error rate for one model at temperature 0.75, so the improvement is not just distribution sharpening.
  • Conventional contrastive decoding is harmful for speech: in the ablation it raises CER, lowers speaker similarity, and compresses utterance duration to roughly two-thirds of native, whereas ECCD's components restore most of that temporal structure.
  • The measured signatures — lower $I_i$ and $G_i$ at first-error boundaries than at matched correct boundaries, with higher values inside complete error regions — imply that alignment support returns during propagation without restoring correct content, which distinguishes onset from propagation.

Reading between the lines

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

  • The text-ablated amateur distribution could be repurposed as an online hallucination-onset detector: tokens where $G_i$ is unusually low relative to a running baseline may pinpoint the transitions where intervention matters most, a use the paper's averaged analysis does not itself establish.
  • The same design pattern — expert anchoring, positive-only contrast, and set-level compatibility calibration — may transfer to other autoregressive generation tasks where the ablated 'amateur' distribution carries useful regularities rather than pure noise, such as image-token or audio-language models.
  • The paper's fixed default of $\alpha=1$, $k=25$ is not optimal in its own hyperparameter sweep, where $\alpha=1$, $k=15$ improves both CER and speaker similarity, so per-model tuning could yield larger gains than the headline numbers.
  • Because $C_i$ measures set-level agreement between experience and expert candidates, it may also serve as a per-step measure of text-history conflict, potentially useful for diagnosing which linguistic units (gaps versus sustained regions) are most vulnerable.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Experience-Calibrated Contrastive Decoding (ECCD), a training-free decoding method for LM-based text-to-speech that contrasts the full-condition next-token distribution p_E against a text-ablated distribution p_A. ECCD retains the original expert score, applies a positive-only alignment enhancement within a top-k plausible set, and scales the enhancement by an experience compatibility coefficient. The authors report WER/CER reductions across four TTS models (CosyVoice2, CosyVoice3, Llasa, GLM-TTS) on SeedTTS-Eval and CV3-Eval, a listening-test CMOS gain of +0.644, and an information-theoretic analysis of alignment influence around hallucination onsets. The paper claims the first adaptation of contrastive decoding to autoregressive acoustic-token generation and interprets the results as evidence that conditional-information control is a promising route for mitigating speech hallucinations.

Significance. If the empirical claims hold, ECCD is a simple, model-agnostic, training-free intervention that could complement existing training-based hallucination mitigations. The evaluation is unusually broad: four models, nine languages, both objective and subjective metrics, a component ablation, and a mechanism analysis. The paper also states its limitations explicitly, including the conditioning-shift caveat for p_A and the approximate nature of the ASR-based onset alignment. The main gap is statistical: the universal 'all settings' claim rests on point estimates without confidence intervals or significance tests, and some reported gains are very small. The mechanism analysis is informative but not fully independent of the proposed method. With additional statistical support and a clearer treatment of the proxy-validity and hyperparameter-selection questions, this would be a valuable contribution.

major comments (4)
  1. [Experiments, Table 1 and Tables 2–3] The headline claim that ECCD 'reduces WER/CER in all SeedTTS-Eval settings and 24 of 25 CV3-Eval settings' is supported only by single point estimates without confidence intervals or significance tests. Several improvements are small (e.g., GLM-TTS test-zh CER 1.01→0.97 and test-en WER 2.12→1.94) and may lie within utterance-level or ASR variability. I recommend reporting bootstrap or per-utterance confidence intervals, or paired significance tests (e.g., McNemar on utterance-level errors), and stating which settings remain significant; this is needed to substantiate the universal claim.
  2. [Method, Eq. (2)] The amateur distribution p_A is obtained by removing all text conditions, but the paper itself notes that 'removing text may itself induce a conditioning shift.' Because both ECCD's enhancement and the I_i/G_i measures are built on the p_E/p_A contrast, the interpretation of the gains—and of the onset analysis—as alignment-specific depends on p_A being a valid experience proxy. I ask for a control experiment or sensitivity analysis, such as comparing p_A against a condition with uninformative text or a model whose text conditioning is separately ablated, to rule out artifacts from altered prosody, silence, or degenerate outputs.
  3. [Experiments, Figure 2 and Table 1] The hyperparameter sweep and the component ablation are run on SeedTTS-Eval test-hard, the same split on which the headline CosyVoice2 results are reported. Although α=1,k=25 is presented as a fixed default, the sweep itself shows α=1,k=15 would give lower CER and higher SS on this split, so the degree to which the reported results benefit from selection on the test set is unquantified. A held-out validation split should be used for any tuning, or the authors should show that the default performs comparably on an independent split.
  4. [Mechanism analysis, Tables 5–6] The I_i and G_i statistics are computed from the same p_E/p_A contrast that ECCD modifies, so the observation that G_i is lower at first-error boundaries than at matched correct boundaries is not an independent test of the proposed onset mechanism; it could reflect a property of the contrast itself (e.g., conditioning-shift artifacts) rather than alignment support. The paper's caveat that the analysis is 'consistent with, but does not causally establish' the hypothesis is welcome, but the section should explicitly state this non-independence and, ideally, support the interpretation with a reference measure not derived from the same contrast (e.g., attention maps or forced alignment).
minor comments (4)
  1. [Equation (8) and ablation of ECCD_I] I_i is an unbounded KL divergence, and the text says it 'replaces' 1−C_i in Eq. (8) without describing normalization; please specify the exact substitution and any normalization used, since this directly affects reproducibility.
  2. [Figure 1] The dashed segments after offset 5 are described as unreliable because sample counts decrease, but no counts or thresholds are given; please include a sample-size profile or a minimum-count note so the reader can judge the reliability of those segments.
  3. [Reproducibility] The paper does not state whether code or audio samples will be released; given the simplicity of the training-free method, providing these would materially help replication.
  4. [General presentation] The extracted text contains frequent missing spaces (e.g., 'acoustictokenselection'), and some model names are used inconsistently; please ensure the camera-ready text is properly spaced and model names are consistent throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the central WER/CER claim is tested against external ASR and listening benchmarks, and the internal mechanism analysis is explicitly operational and hedged.

full rationale

The paper's central claim is that ECCD reduces hallucination-related content errors across four TTS models. This is evaluated with external ASR systems (Whisper-large-v3, Paraformer-zh) and a human listening test with CMOS, so the main result is not derived from the method's own definitions or fitted to the same data. The ECCD score in Eq. (8) is a proposed decoding rule, not a mathematical consequence of the alignment measures in Eqs. (4)/(5); its effectiveness is established empirically. The only potentially self-referential element is the secondary information-theoretic analysis: I_i and G_i are computed from the same full-condition versus text-ablated contrast that ECCD uses, so the analysis cannot independently confirm the causal mechanism. However, the paper explicitly labels p_A an 'operational proxy' and states that 'this comparison is not an exact decomposition of alignment and experience information' and that removing text 'may itself induce a conditioning shift.' The onset analysis is also carefully hedged: the lower values at first-error boundaries are described as 'consistent with, but do not causally establish, an alignment-support shortfall.' No parameter is fitted and then renamed a prediction: alpha=1 and k=25 are fixed defaults, and the hyperparameter sweep explicitly notes that alpha=1, k=15 would be a favorable deployment setting when validation-based tuning is available. The single self-citation to Liu et al. (2025) is used only to support the observation that hallucinated speech can remain fluent, which is also independently plausible and not the load-bearing premise of the method. Lack of confidence intervals for WER/CER is a statistical rigor concern, not circularity. Overall, the derivation chain does not reduce to its inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on two free hyperparameters (alpha, k), the operational proxy pA for experience information, the top-k plausible set, and the ASR-based token-span approximation. The paper introduces no new physical entities; the alignment/experience distinction is an operational labeling of existing distributions, not an independently evidenced entity.

free parameters (2)
  • alpha (alignment-enhancement scale) = 1 (default; swept over 0.5, 1, 2, 4)
    Scales the positive alignment enhancement in Eq. (8). Chosen by hand and via hyperparameter sweep, not derived.
  • k (plausible set size) = 25 (default; swept over 15, 50, 100)
    Top-k expert-defined plausible set size in Eq. (6). Chosen by hand and via hyperparameter sweep, not derived.
assumptions (4)
  • domain assumption The text-ablated distribution pA(x_i | x_<i, a) is an operational proxy for experience-oriented prediction.
    Eq. (2). The paper states this is 'not an exact decomposition' and that removing text may induce a conditioning shift, yet the method and the I_i/G_i analysis rely on this proxy.
  • domain assumption The top-k expert-defined plausible set V_head is a valid constraint on admissible tokens.
    Eq. (6). Used to restrict enhancement to expert-plausible candidates; follows standard CD practice.
  • domain assumption ASR transcripts, timestamps, and the model token rate approximate character-to-token alignment.
    Analysis setup. The authors acknowledge these windows localize the observed transition rather than the exact causal decoding step.
  • standard math Autoregressive factorization p(x | a, T_p, T_t) = product_i p(x_i | x_<i, a, T_p, T_t).
    Eq. (1). Standard for autoregressive LMs and uncontroversial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Experience-Calibrated Contrastive Decoding for Mitigating Hallucinations in LM-Based Text-to-Speech." pith.science (2026). https://pith.science/paper/LWBJZFRI

@misc{pith2026260800722,
  author       = {Pith},
  title        = {Pith review of: Experience-Calibrated Contrastive Decoding for Mitigating Hallucinations in LM-Based Text-to-Speech},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LWBJZFRI}},
  note         = {Machine review of arXiv:2608.00722}
}
abstract

Language model-based text-to-speech (LM-based TTS) remains vulnerable to speech hallucinations that deviate from the target text. Existing mitigation mainly relies on architectural changes or additional training, while decoding-time control remains underexplored. We present a conditional information view that distinguishes text-derived alignment information from experience information supplied by acoustic context and learned speech regularities. We hypothesize that an important class of hallucinations begins when alignment support is insufficiently reflected in the selected token at a vulnerable transition. Using predictions from the same speech LM with and without text conditions, we propose Experience-Calibrated Contrastive Decoding (ECCD), a training-free method that strengthens alignment support while preserving useful experience information. ECCD preserves the original expert distribution, applies only positive alignment enhancement, and calibrates its strength using set-level experience compatibility. Across four models, ECCD reduces WER/CER by up to 55.6% in all SeedTTS-Eval settings and 24 of 25 multilingual CV3-Eval settings. A listening test yields a CMOS gain of $+0.644$ while retaining strong speaker similarity. Further analysis shows that alignment influence and decision-level gain vary within linguistic units and are lower at first-error boundaries than at matched correct boundaries. Overall, these extensive experiments and analyses identify conditional information control as a promising decoding-time direction for mitigating speech hallucination.

Figures

Figures reproduced from arXiv: 2608.00722 by the authors.

Figure 1
Figure 1. Mean relative-position profiles of Ii (top) and Gi (bottom) over ASR-aligned correct and erroneous character regions. Offset 0 is the first token. Segments after offset 5 are dashed because their sample counts decrease sharply. bration signal. To quantify temporal compression, we report two descriptive statistics: utterance-level average acoustic￾token count (UAC) and timestamp-derived average character duration (TA… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

115 extracted references · 29 canonical work pages

  1. [1]

    Artificial Intelligence Review , volume=

    Conventional and contemporary approaches used in text to speech synthesis: A review , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

    2023

  2. [2]

    Advances in neural information processing systems , volume=

    Bridging the gap between value and policy based reinforcement learning , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2307.09288 , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    arXiv

  4. [4]

    Proceedings of the 24th international conference on Machine learning , pages=

    Reinforcement learning by reward-weighted regression for operational space control , author=. Proceedings of the 24th international conference on Machine learning , pages=

  5. [5]

    International conference on machine learning , pages=

    Reinforcement learning with deep energy-based policies , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  6. [6]

    Multimedia Tools and Applications , volume=

    A deep learning approaches in text-to-speech system: a systematic review and recent research perspective , author=. Multimedia Tools and Applications , volume=. 2023 , publisher=

    2023

  7. [7]

    Interspeech 2017 , pages=

    Tacotron: Towards End-to-End Speech Synthesis , author=. Interspeech 2017 , pages=. 2017 , publisher=

    2017

  8. [8]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Contrastive decoding: Open-ended text generation as optimization , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

Show all 115 references
  1. [9]

    Nature , volume=

    DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning , author=. Nature , volume=. 2025 , publisher=

    2025

  2. [10]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    arXiv

  3. [11]

    ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Group relative policy optimization for text-to-speech with large language models , author=. ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2026 , organization=

    2026

  4. [12]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  5. [13]

    Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    Sequence-level knowledge distillation , author=. Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    2016

  6. [14]

    arXiv preprint arXiv:2308.01825 , year=

    Scaling relationship on learning mathematical reasoning with large language models , author=. arXiv preprint arXiv:2308.01825 , year=

    arXiv

  7. [15]

    Neural Information Processing Systems , year=

    The Llama 3 herd of models , author=. Neural Information Processing Systems , year=

  8. [16]

    arXiv preprint arXiv:2601.15621 , year=

    Qwen3-TTS Technical Report , author=. arXiv preprint arXiv:2601.15621 , year=

    arXiv

  9. [17]

    International Conference on Learning Representations , year=

    The Curious Case of Neural Text Degeneration , author=. International Conference on Learning Representations , year=

  10. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Ella-v: Stable neural codec language modeling with alignment-guided sequence reordering , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  11. [19]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Hierarchical neural story generation , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  12. [20]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  13. [21]

    The Twelfth International Conference on Learning Representations , year=

    SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models , author=. The Twelfth International Conference on Learning Representations , year=

  14. [22]

    arXiv preprint arXiv:2406.17310 , year=

    High fidelity text-to-speech via discrete tokens using token transducer and group masked language model , author=. arXiv preprint arXiv:2406.17310 , year=

    arXiv

  15. [23]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Repcodec: A speech representation codec for speech tokenization , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  16. [24]

    Deep voice 3: 2000-speaker neural text-to-speech , author=. proc. ICLR , volume=

    2000

  17. [25]

    Advances in neural information processing systems , volume=

    Fastspeech: Fast, robust and controllable text to speech , author=. Advances in neural information processing systems , volume=

  18. [26]

    ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech , author=

  19. [27]

    Speech Synthesis Based on Hidden Markov Models , year=

    Tokuda, Keiichi and Nankaku, Yoshihiko and Toda, Tomoki and Zen, Heiga and Yamagishi, Junichi and Oura, Keiichiro , journal=. Speech Synthesis Based on Hidden Markov Models , year=

  20. [28]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Neural speech synthesis with transformer network , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  21. [29]

    FastSpeech 2: Fast and High-Quality End-to-End Text to Speech , author=

  22. [30]

    Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers , year=

    Chen, Sanyuan and Wang, Chengyi and Wu, Yu and Zhang, Ziqiang and Zhou, Long and Liu, Shujie and Chen, Zhuo and Liu, Yanqing and Wang, Huaming and Li, Jinyu and He, Lei and Zhao, Sheng and Wei, Furu , journal=. Neural Codec Language Models are Zero-Shot Text to Speech Synthesi...

  23. [31]

    Transactions of the Association for Computational Linguistics , volume=

    Speak, read and prompt: High-fidelity text-to-speech with minimal supervision , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher=

    2023

  24. [32]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Voicecraft: Zero-shot speech editing and text-to-speech in the wild , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  25. [33]

    arXiv preprint arXiv:2512.14291 , year=

    Glm-tts technical report , author=. arXiv preprint arXiv:2512.14291 , year=

  26. [34]

    Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation , year=

    VALL-E R: Robust and Efficient Zero-Shot Text-to-Speech Synthesis via Monotonic Alignment , author=. Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation , year=

    2024

  27. [35]

    arXiv preprint arXiv:2407.05407 , year=

    Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens , author=. arXiv preprint arXiv:2407.05407 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2412.10117 , year=

    CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models , author=. arXiv preprint arXiv:2412.10117 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2505.17589 , year=

    Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training , author=. arXiv preprint arXiv:2505.17589 , year=

    arXiv

  30. [38]

    Interspeech 2019 , year=

    LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech , author=. Interspeech 2019 , year=

    2019

  31. [39]

    WenetSpeech4TTS: A 12,800-hour Mandarin TTS Corpus for Large Speech Generation Model Benchmark , author=. Proc. Interspeech 2024 , pages=

    2024

  32. [40]

    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition , author=. Proc. Interspeech 2022 , pages=

    2022

  33. [41]

    International conference on machine learning , pages=

    Robust speech recognition via large-scale weak supervision , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  34. [42]

    CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking , author=. Proc. Interspeech 2023 , pages=

    2023

  35. [43]

    Interspeech 2022 , year=

    UTMOS: UTokyo-SaruLab System for VoiceMOS Challenge 2022 , author=. Interspeech 2022 , year=

    2022

  36. [44]

    arXiv preprint arXiv:2406.02430 , year=

    Seed-tts: A family of high-quality versatile speech generation models , author=. arXiv preprint arXiv:2406.02430 , year=

    arXiv

  37. [45]

    ACM computing surveys , volume=

    Survey of hallucination in natural language generation , author=. ACM computing surveys , volume=. 2023 , publisher=

    2023

  38. [46]

    arXiv preprint arXiv:1606.06565 , year=

    Concrete problems in AI safety , author=. arXiv preprint arXiv:1606.06565 , year=

    arXiv

  39. [47]

    International Conference on Learning Representations , year=

    The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models , author=. International Conference on Learning Representations , year=

  40. [48]

    The Twelfth International Conference on Learning Representations , year=

    DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models , author=. The Twelfth International Conference on Learning Representations , year=

  41. [49]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    NaturalSpeech 3: zero-shot speech synthesis with factorized codec and diffusion models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  42. [50]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Enabling beam search for language model-based text-to-speech synthesis , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  43. [51]

    arXiv preprint arXiv:2502.04128 , year=

    Llasa: Scaling Train-Time and Inference-Time Compute for Llama-based Speech Synthesis , author=. arXiv preprint arXiv:2502.04128 , year=

    arXiv

  44. [52]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  45. [53]

    IEEE/ACM transactions on audio, speech, and language processing , volume=

    Audiolm: a language modeling approach to audio generation , author=. IEEE/ACM transactions on audio, speech, and language processing , volume=. 2023 , publisher=

    2023

  46. [54]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Autoregressive speech synthesis without vector quantization , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  47. [55]

    arXiv preprint arXiv:2412.16846 , year=

    Autoregressive Speech Synthesis with Next-Distribution Prediction , author=. arXiv preprint arXiv:2412.16846 , year=

  48. [56]

    arXiv preprint arXiv:2212.08073 , year=

    Constitutional ai: Harmlessness from ai feedback , author=. arXiv preprint arXiv:2212.08073 , year=

    arXiv

  49. [57]

    Advances in Neural Information Processing Systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in Neural Information Processing Systems , volume=

  50. [58]

    arXiv preprint arXiv:2402.10571 , year=

    Direct preference optimization with an offset , author=. arXiv preprint arXiv:2402.10571 , year=

    arXiv

  51. [59]

    The Twelfth International Conference on Learning Representations , year=

    Amortizing intractable inference in large language models , author=. The Twelfth International Conference on Learning Representations , year=

  52. [60]

    International Conference on Machine Learning , pages=

    Scaling laws for reward model overoptimization , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  53. [61]

    Advances in Neural Information Processing Systems , volume=

    Flow network based generative models for non-iterative diverse candidate generation , author=. Advances in Neural Information Processing Systems , volume=

  54. [62]

    The Journal of Machine Learning Research , volume=

    Gflownet foundations , author=. The Journal of Machine Learning Research , volume=. 2023 , publisher=

    2023

  55. [63]

    arXiv preprint arXiv:2406.00633 , year=

    Improving GFlowNets for Text-to-Image Diffusion Alignment , author=. arXiv preprint arXiv:2406.00633 , year=

    arXiv

  56. [64]

    arXiv preprint arXiv:2409.05746 , year=

    Llms will always hallucinate, and we need to live with this , author=. arXiv preprint arXiv:2409.05746 , year=

    arXiv

  57. [65]

    arXiv preprint arXiv:2406.00654 , year=

    Enhancing zero-shot text-to-speech synthesis with human feedback , author=. arXiv preprint arXiv:2406.00654 , year=

    arXiv

  58. [66]

    arXiv preprint arXiv:2410.15326 , year=

    A survey of uncertainty estimation in llms: Theory meets practice , author=. arXiv preprint arXiv:2410.15326 , year=

    arXiv

  59. [67]

    arXiv preprint arXiv:2502.00290 , year=

    Estimating LLM Uncertainty with Logits , author=. arXiv preprint arXiv:2502.00290 , year=

    arXiv

  60. [68]

    ACM Computing Surveys , volume=

    A survey on uncertainty quantification of large language models: Taxonomy, open research challenges, and future directions , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  61. [69]

    Advances in Neural Information Processing Systems , volume=

    To believe or not to believe your llm: Iterative prompting for estimating epistemic uncertainty , author=. Advances in Neural Information Processing Systems , volume=

  62. [70]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Enhancing uncertainty-based hallucination detection with stronger focus , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  63. [71]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  64. [72]

    arXiv preprint arXiv:2407.06426 , year=

    DebUnc: mitigating hallucinations in large language model agent communication with uncertainty estimations , author=. arXiv preprint arXiv:2407.06426 , year=

    arXiv

  65. [73]

    arXiv preprint arXiv:2303.03926 , year=

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling , author=. arXiv preprint arXiv:2303.03926 , year=

    arXiv

  66. [74]

    arXiv preprint arXiv:2406.05370 , year=

    Vall-e 2: Neural codec language models are human parity zero-shot text to speech synthesizers , author=. arXiv preprint arXiv:2406.05370 , year=

    arXiv

  67. [75]

    arXiv preprint arXiv:2406.04904 , year=

    Xtts: a massively multilingual zero-shot text-to-speech model , author=. arXiv preprint arXiv:2406.04904 , year=

    arXiv

  68. [76]

    arXiv preprint arXiv:2402.08093 , year=

    Base tts: Lessons from building a billion-parameter text-to-speech model on 100k hours of data , author=. arXiv preprint arXiv:2402.08093 , year=

    arXiv

  69. [77]

    arXiv preprint arXiv:2412.08237 , year=

    TouchTTS: An Embarrassingly Simple TTS Framework that Everyone Can Touch , author=. arXiv preprint arXiv:2412.08237 , year=

    arXiv

  70. [78]

    arXiv preprint arXiv:2409.18512 , year=

    EmoPro: A Prompt Selection Strategy for Emotional Expression in LM-based Speech Synthesis , author=. arXiv preprint arXiv:2409.18512 , year=

    arXiv

  71. [79]

    arXiv preprint arXiv:2503.01710 , year=

    Spark-tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens , author=. arXiv preprint arXiv:2503.01710 , year=

    arXiv

  72. [80]

    Advances in Neural Information Processing Systems , volume=

    Trajectory balance: Improved credit assignment in gflownets , author=. Advances in Neural Information Processing Systems , volume=

  73. [81]

    International Conference on Machine Learning , pages=

    Learning gflownets from partial episodes for improved convergence and stability , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  74. [82]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  75. [83]

    International Conference on Machine Learning , pages=

    Biological sequence design with gflownets , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  76. [84]

    Uncertainty in Artificial Intelligence , pages=

    Bayesian structure learning with generative flow networks , author=. Uncertainty in Artificial Intelligence , pages=. 2022 , organization=

    2022

  77. [85]

    International Conference on Machine Learning , pages=

    Generative flow networks for discrete probabilistic modeling , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  78. [86]

    arXiv preprint arXiv:2209.02606 , year=

    Unifying generative models with GFlowNets and beyond , author=. arXiv preprint arXiv:2209.02606 , year=

    arXiv

  79. [87]

    International Conference on Machine Learning , pages=

    Better training of gflownets with local credit and incomplete trajectories , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  80. [88]

    International Conference on Artificial Intelligence and Statistics , pages=

    Generative flow networks as entropy-regularized rl , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

    2024

  81. [89]

    International Conference on Machine Learning , pages=

    A theory of continuous generative flow networks , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  82. [90]

    Transactions on Machine Learning Research , year=

    A Variational Perspective on Generative Flow Networks , author=. Transactions on Machine Learning Research , year=

  83. [91]

    The Eleventh International Conference on Learning Representations , year=

    GFlowNets and variational inference , author=. The Eleventh International Conference on Learning Representations , year=

  84. [92]

    Advances in neural information processing systems , volume=

    Compositional sculpting of iterative generative processes , author=. Advances in neural information processing systems , volume=

  85. [93]

    The Eleventh International Conference on Learning Representations , year=

    Generative Augmented Flow Networks , author=. The Eleventh International Conference on Learning Representations , year=

  86. [94]

    Digital Discovery , volume=

    Gflownets for ai-driven scientific discovery , author=. Digital Discovery , volume=. 2023 , publisher=

    2023

  87. [95]

    Advances in Neural Information Processing Systems , volume=

    Rgfn: Synthesizable molecular generation using gflownets , author=. Advances in Neural Information Processing Systems , volume=

  88. [96]

    NeurIPS 2023 Generative AI and Biology (GenBio) Workshop , year=

    Generative flow networks assisted biological sequence editing , author=. NeurIPS 2023 Generative AI and Biology (GenBio) Workshop , year=

    2023

  89. [97]

    Advances in neural information processing systems , volume=

    Let the flows tell: Solving graph combinatorial problems with gflownets , author=. Advances in neural information processing systems , volume=

  90. [98]

    The Eleventh International Conference on Learning Representations , year=

    Robust Scheduling with GFlowNets , author=. The Eleventh International Conference on Learning Representations , year=

  91. [99]

    arXiv preprint arXiv:2403.07041 , year=

    Ant colony sampling with gflownets for combinatorial optimization , author=. arXiv preprint arXiv:2403.07041 , year=

    arXiv

  92. [100]

    Proceedings of the 38th International Conference on Neural Information Processing Systems , pages=

    Amortizing intractable inference in diffusion models for vision, language, and control , author=. Proceedings of the 38th International Conference on Neural Information Processing Systems , pages=

  93. [101]

    arXiv preprint arXiv:2412.07775 , year=

    Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets , author=. arXiv preprint arXiv:2412.07775 , year=

    arXiv

  94. [102]

    Proceedings of the 31st ACM International Conference on Multimedia , pages=

    Generalized universal domain adaptation with generative flow networks , author=. Proceedings of the 31st ACM International Conference on Multimedia , pages=

  95. [103]

    arXiv preprint arXiv:2310.08774 , year=

    PhyloGFN: Phylogenetic inference with generative flow networks , author=. arXiv preprint arXiv:2310.08774 , year=

    arXiv

  96. [104]

    arXiv preprint arXiv:2509.19852 , year=

    Eliminating stability hallucinations in llm-based tts models via attention guidance , author=. arXiv preprint arXiv:2509.19852 , year=

  97. [105]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Mitigating hallucinations in lm-based tts models via distribution alignment using gflownets , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  98. [106]

    Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding , year=

    Hsu, Tzu-Wen and Lu, Ke-Han and Chiang, Cheng-Han and Lee, Hung-Yi , booktitle=. Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding , year=

  99. [107]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    Temporal Contrastive Decoding: A Training-Free Method for Large Audio-Language Models , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  100. [108]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Vall-t: Decoder-only generative transducer for robust and decoding-controllable text-to-speech , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  101. [109]

    2024 IEEE Spoken Language Technology Workshop (SLT) , pages=

    Attention-constrained inference for robust decoder-only text-to-speech , author=. 2024 IEEE Spoken Language Technology Workshop (SLT) , pages=. 2024 , organization=

    2024

  102. [110]

    doi:10.21437/Interspeech.2024-335 , issn =

    Paarth Neekhara and Shehzeen Hussain and Subhankar Ghosh and Jason Li and Boris Ginsburg , year =. doi:10.21437/Interspeech.2024-335 , issn =

    2024 doi

  103. [111]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mitigating object hallucinations in large vision-language models through visual contrastive decoding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  104. [112]

    International Conference on Learning Representations , volume=

    Dola: Decoding by contrasting layers improves factuality in large language models , author=. International Conference on Learning Representations , volume=

  105. [113]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Contrastive decoding reduces hallucinations in large multilingual machine translation models , author=. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  106. [114]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Evidential-TTS: High Fidelity Zero-Shot Text-to-Speech Using Evidential Deep Learning , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  107. [115]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    Trusting your evidence: Hallucinate less with context-aware decoding , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.