WorksComputer Science
LoCA: Local Credit Assignment for Reinforcement Learning of Autoregressive Flow-Matching TTS
Abstract
Reinforcement learning can improve text-to-speech synthesis through feedback on generated speech. However, sentence-level rewards do not identify local errors, while signed regression updates can lead to excessive fitting or repulsion of sampled outputs. We propose LoCA (Local Credit Assignment) for autoregressive flow-matching text-to-speech. LoCA uses ASR-based text alignment to locate content errors in generated speech and redistribute sentence-level advantages across acoustic frames. To improve training stability, we introduce reference-relative positive and negative thresholds that limit continued fitting and repulsion. Experiments on VoxCPM2 show that LoCA reduces sample-weighted content error on a multilingual CV3 evaluation subset by 20.5% and 14.3% relative to the pretrained backbone and GROW, respectively, while improving speaker similarity. LoCA also improves content accuracy on the Seed-TTS-Eval hard subset. Ablation results support the contributions of local credit and threshold-based regression control.
Introduction
Speech synthesis still produces mispronunciations, omissions and repetitions, motivating post-training with generated-speech feedback [1, 5]. Continuous-latent models such as DiTAR [11] and the VoxCPM family [24, 23] combine autoregressive context with diffusion or flow-based acoustics; their regression objectives provide an interface for reward-driven optimization.
This setting raises two questions. First, where should a sentence reward act? A single mispronounced word or silent interval may account for an utterance’s low score, while broadcasting one coefficient treats that span and its surrounding speech identically. Second, when should a regression update stop? Positive advantages encourage fitting sampled trajectories, whereas negative advantages encourage increasing their regression error. A reference anchor opposes drift but does not specify separate stopping conditions for fitting and repulsion.
We propose LoCA (Local Credit Assignment) to address both questions. LoCA centers available frame labels within each utterance and adds them to its sentence advantage. This redistributes learning coefficients while preserving their utterance mean. For the second, we introduce reference-relative positive and negative thresholds. An improvement budget limits continued fitting, an upper error threshold stops continued repulsion, and the reference penalty remains active. Together, these components enable LoCA to target higher synthesis accuracy and more stable training without changing inference.
Evaluation should also distinguish average accuracy from severe failures. Mean CER/WER alone does not reveal how often synthesis produces silence, major content errors or truncated outputs. We define an automatic severe failure rate (SFR) using explicit content-error, silence and generation-failure conditions. An utterance satisfying any condition is counted once; SFR is the proportion of flagged utterances in the full evaluation set. We use SFR alongside CER/WER as a core metric for comparing synthesis performance across methods and training stages.
Our contributions are local credit for targeted content learning and two-sided regression control for quality improvement and reduced drift. Baseline comparisons and component ablations on VoxCPM2 examine both components using hard-set trajectories, selected multilingual CV3 splits and standard Seed evaluation.
Related Work
Speech post-training includes multi-reward optimization in Seed-TTS ([1]), GRPO in GLM-TTS ([5]), differentiable reward optimization in DiffRO ([7]), and token-level preference learning in TKTO ([12]). These methods motivate targeted learning from synthesis feedback; our setting uses continuous acoustic regression rather than discrete-token policy updates.
GROW ([21]) directly weights flow-matching regression with group-relative sentence advantages and anchors velocities to a frozen reference. FlowAWR ([6]), evaluated in image generation, derives an advantage-rectified velocity target from a KL-constrained reward objective; AdvantageFlow ([13]) regularizes advantage-weighted regression through the rollout policy. We build on the direct-regression approach, introducing within-utterance credit and explicit positive/negative stopping thresholds instead of changing the velocity target. GROW’s existing reference anchor is retained.
LoCA: Local Credit with Two-Sided Control
From sentence feedback to signed regression
For each text and reference-voice condition, generate candidates, score their audio, and retain their latents for replay. Let be the content reward and speaker similarity. Using group means and sample standard deviations, define
Separate normalization controls the reward scales; constant components contribute zero.
For generated latents and Gaussian noise , the data-at-zero flow path (Lipman et al. [14]) is
The policy velocity and frozen pretrained reference receive the same generated history, noisy input and flow time. For channel dimension , let
Following GROW (Yang et al. [21]), the sentence-weighted starting point is
where counts generated frames. Its reward term has derivative with respect to before averaging: the sign selects fitting or repulsion, but every frame of an utterance shares that sign and coefficient.
Constructing and applying local credit
From text errors to frame labels. We transcribe each generated utterance with Qwen3-ASR (Shi et al. [20]) and align the recognized words to its waveform with the forced aligner. Edit alignment against the target text identifies substitutions and insertions; their timestamps locate error spans. Deletions use neighboring-word boundaries because omitted words have no emitted span. These intervals are projected onto the exact generated latent-frame timeline, rather than treated as acoustic tokens.
Availability-aware feedback. Each frame receives : unknown, detected error, or unflagged. Unknown includes unreliable timing and the unmeasured tail; it is not a negative reward. Confirmed anomalous silence is labeled 0, while eligible exact-phonetic substitutions become unknown. Define . We center only measured labels within each complete utterance:
Error and unflagged frames receive relative negative and positive increments; unknown frames receive zero. The increments sum to zero, preserving the sentence coefficient’s mean. Uniform or wholly unknown labels yield zero increments. Centering is per utterance, not across candidates or micro-batches.
Integration with regression. We broadcast to the generated frames and replace it by in the regression objective below. Labels and coefficients are detached; prefix, padding and stop positions are excluded. No separate local reward loss is added. The final coefficient need not change sign, but can do so when local evidence outweighs the sentence advantage (Figure 1). Appendix § 7 specifies timing, deletion and label-priority rules.
Figure 1. Local-credit construction in an illustrative 12-frame example. Six unflagged and four error frames give mean label 0.6; two unknown tail frames are excluded. With local weight 0.2 and sentence advantage 0.05, error frames receive final coefficient -0.07, unflagged frames +0.13, and unknown frames retain +0.05. The boxes represent generated latent frames, not discrete speech tokens.
Two-sided thresholds for controlled updates
We next control how long either sign of regression remains active. With positive budget , negative margin and relative fraction , define
The relative budget keeps the positive floor nonnegative even when reference error is small. Using , the combined objective is
Feedback, targets and reference predictions are detached. Each active branch retains the signed-regression derivative: positive fitting stops at the improvement budget, and negative repulsion stops at the upper threshold. The reference penalty continues to anchor both branches. This removes the incentive for unbounded negative regression and limits repeated positive fitting. Appendix B.1 derives the gradients and fixed-replay bounds; these properties concern the regression objective rather than a global bound on parameter drift.
Experiments
Backbone, data and evaluation
All systems use the publicly released 2B-parameter VoxCPM2 checkpoint (openbmb/VoxCPM2; [16]). The base is evaluated without RL; GROW, LoCA and the ablations start from this same checkpoint.
RL samples a 50,000-example Emilia ([9]) pool covering Chinese, English, German, French, Japanese and Korean. Hyperparameters are in Appendix § 6. Evaluation uses seven CV3 ([17]) language splits (Chinese, English, German, Spanish, French, Italian and Russian; 500 each), plus hard Chinese/English (60/64), totaling 3,624. This is a fixed subset of CV3. Seed-TTS-Eval ([2]) includes Chinese hard (400) and standard Chinese/English (2,020/1,088).
We select representative models by jointly considering content accuracy, severe failures and speaker similarity. Complete training trajectories and supplementary results are provided in Appendix § 8.
Automatic severe failure rate
For item , let be utterance CER/WER; , and indicate all-silent output, a generation-cap hit and a generation error. Let and be tail and longest internal silence in seconds:
The denominator is the full evaluation set, including generation errors. Each flagged item counts once; tables and plots report this union count as severe proxy cases. Detection details are in Appendix § 9.
Main results and ablations
LoCA achieves CV3 error 4.758%, versus 5.986% for the base and 5.553% for GROW (Table 1). Severe proxy cases decrease from 61/60 to 51, and SIM rises from 67.168/67.562 to 69.137. On ordinary CV3, LoCA has error 4.673%, versus 5.887%/5.461%. It also lowers Seed-hard CER to 6.884% with 5 severe cases. CV3-hard error is 7.163%, with 6 cases: lower error than both baselines, with the base’s case count and one fewer case than GROW.
| Method | CV3 selected (3,624) Error ↓ | CV3 selected (3,624) Cases ↓ | CV3 selected (3,624) SIM ↑ | CV3 regular Error ↓ | Seed hard Error ↓ | Seed hard Cases ↓ | CV3 hard Error ↓ | CV3 hard Cases ↓ |
| VoxCPM2 base | 5.986 | 61 | 67.168 | 5.887 | 8.193 | 6 | 8.772 | 6 |
| GROW | 5.553 | 60 | 67.562 | 5.461 | 9.575 | 9 | 8.157 | 7 |
| LoCA | 4.758 | 51 | 69.137 | 4.673 | 6.884 | 5 | 7.163 | 6 |
Table 1. Main model comparison. Every metric in a row uses the same selected model. Cases are severe proxy counts and SIM is cosine similarity scaled by 100. Lower Error/Cases and higher SIM are preferred. CV3 regular covers the 3,500 ordinary-language items; the full selection also includes 124 hard items. CV3 hard weights 60 Chinese and 64 English items.
Table 2 evaluates component removals. Without local credit, CV3 error is 5.508%, with 54 cases and SIM 67.525. Without both thresholds, it is 5.488%, with 44 cases and SIM 67.722. Full LoCA has better content accuracy and speaker similarity than both variants; the no-threshold variant has fewer severe cases. The trajectories retain these metric-dependent trade-offs.
| LoCA variant | Local | Thresholds | Error ↓ | Cases ↓ | SIM ↑ |
| LoCA (full) | Yes | Yes | 4.758 | 51 | 69.137 |
| w/o local credit | No | Yes | 5.508 | 54 | 67.525 |
| w/o thresholds | Yes | No | 5.488 | 44 | 67.722 |
Table 2. LoCA component ablations. Yes/No indicates component presence. Metrics use the full 3,624-item CV3 selection. The variant w/o thresholds removes the positive and negative regression thresholds.
Russian generalization
Russian is excluded from the six-language RL pool. On its 500-item CV3 split, LoCA’s selected checkpoint reduces WER from the base’s 6.219% to 4.723% (24.1% relative) and raises SIM from 67.296 to 68.541; severe proxy cases are 6 versus the base’s 5. WER remains below the base at every evaluated LoCA checkpoint. This supports transfer beyond the RL training languages; it does not imply Russian was absent from backbone pretraining. Language profiles and similarity trajectories are in Appendix § 10.
Conclusion
LoCA combines local credit with two-sided regression control for continuous TTS. The selected model improves CV3 and Seed-hard content accuracy over the baselines, with higher CV3 speaker similarity. Russian results show content and similarity gains outside the RL training pool. Full trajectories retain the cross-checkpoint and hard-set trade-offs.
Limitations
Experiments use one training seed and one backbone. Results describe the evaluated language and task subsets.
Local labels inherit recognition and timing errors. Phonetic matching is a text-side proxy, while the severe-failure metric is an automatic diagnostic. Fixed-replay threshold properties do not guarantee optimizer convergence or absence of parameter drift.
Ethical Considerations
Voice cloning can enable impersonation. Audio/model release requires appropriate dataset permissions and speaker rights; this study does not es tablish redistribution rights for every source recording. Automatic error measures do not guarantee fairness across accents or dialects.
Training and Evaluation Protocol
Training. All trained configurations start from the BF16 release of the 2B-parameter VoxCPM2 model ([23, 16]); GROW denotes its advantage-weighted objective applied to this backbone. The 50,000-example Emilia pool spans Chinese, English, German, French, Japanese and Korean. Pool size is distinct from training duration, which is specified by steps. We use group size 8, replay microbatches of two, and two-condition gradient accumulation. AdamW ([15]) uses learning rate , 50-step warmup, zero weight decay, gradient clipping at 0.5 and seed 42. LoCA uses , , and .
Training samples use ten flow steps, classifier-free guidance ([10]) scale 2 and a 20-second duration ceiling. Generated latents are retained for replay. The audio autoencoder, patch encoder and stop head are frozen; no supervised fine-tuning or stop loss is added. Losses are averaged over generated frames. Content rewards use Qwen3-ASR ([20]); speaker rewards use similarity to the reference voice.
Evaluation. All systems use reference-audio conditioning, ten flow steps, guidance scale 2 and fixed per-example seeds, without retries or best-of- selection. CV3 includes only the seven ordinary splits and two hard splits listed in the main text; it excludes Japanese/Korean and other CV3 tasks. Russian, Spanish and Italian are outside the RL training language pool.
Chinese uses a Paraformer-family recognizer ([8]); other evaluated languages use Whisper large-v3 ([18]). Content error is the untruncated mean of utterance CER/WER. CV3 aggregates these rates by sample count. SIM is speaker-embedding cosine similarity times 100, using ERes2Net for CV3 ([4]) and the benchmark’s fine-tuned WavLM verifier for Seed ([3, 2]). DNSMOS-P808 is an auxiliary automatic quality estimate ([19]), not human MOS.
Local Labels and Objective Properties
The alignment procedure compares normalized reference and ASR transcripts, then uses timestamps aligned to the ASR hypothesis. For waveform duration and generated latent frames, an error interval from to maps to frame indices from to (end exclusive), clipped to the generated range. The mask contains exactly generated-frame entries.
Measured regions start at label 1; substitutions and insertions overwrite their aligned spans with 0. Interior deletions mark the boundary between adjacent recognized tokens, using an 80-ms window when no gap exists. Beginning/end deletions use an 80-ms window clipped to the neighboring word. These are localization heuristics, not timestamps of an emitted missing word. Frames beyond the final reliable timestamp are unknown (). Invalid alignment, near-silent audio or generation-cap hits disable ASR labels; independently confirmed silence can still supply local evidence.
Silence labels require an anomalous-silence detection supported by waveform energy. A 0.3-second grace interval avoids penalizing uncertain boundaries. Only frames fully inside the retained silence interval are marked 0. These acoustic labels override ASR labels, including phonetic-neutral regions. The recognition/timestamp helpers build on public GLM-TTS code ([22]); the within-utterance centered construction is the method studied here.
Phonetic neutralization applies only to equal-length substitution chunks with nonempty, exactly matching pronunciation sequences and valid timestamps. Eligible error labels become unavailable, preserving overlapping insertion/deletion penalties and acoustic overlays. Uniform all-error labels have no within-utterance contrast and hence zero local increment; sentence feedback may still update those candidates.
Fixed-replay derivations
Mean-preserving local credit. For an utterance with , substituting Eq. (5) gives
When , every increment is zero by the mask. If the fraction of unflagged measured frames is , error and unflagged frames receive increments and , respectively. This preserves the mean coefficient before thresholding; different residuals and network Jacobians prevent interpreting it as gradient conservation.
Why an anchor alone need not bound signed regression. Fix one replay frame, its coefficient , target and reference . Expanding the unaveraged signed loss yields
where is independent of . If , the quadratic tends to negative infinity as the output norm grows. This is a sufficient condition for a problematic frame objective, not a claim that every GROW update diverges.
Thresholded derivatives. Let be the unaveraged RL term. Away from the hinge boundaries,
Hence each active branch matches signed regression, and its RL gradient vanishes after saturation. Because , the active negative branch satisfies
All terms of the thresholded loss are nonnegative. For , the per-frame total also satisfies , so it grows as the output moves arbitrarily far from the fixed reference. The positive branch is not upper-bounded in error. These statements concern a fixed replay output; the hinge is nonsmooth at its boundary, and shared parameters can move a saturated frame again. They establish control of the regression terms, not optimizer convergence or globally drift-free training.
Checkpoint Trajectories and Compact Summaries
Figure 2 retains the hard-set training trajectories through step 2000. CV3 hard combines the 60 Chinese and 64 English items: content error and SIM use sample weights, whereas severe case counts are summed. Each group ends at its last evaluated point; in particular, the no-local CV3 series ends at 1800 while its Seed-hard series reaches 2000.
Figure 2. Hard-set trajectories for the baselines and LoCA ablations. Columns show content error, severe proxy cases and speaker similarity. Chinese/English CV3 hard means use weights 60/124 and 64/124; case counts are summed. The dashed line is the base. Lines connect evaluated points without smoothing or extrapolation.
Each model’s main-table metrics, language scores and radar profile use the same selected checkpoint. The trajectories show how all three metrics vary during training.
Standard Seed has a separate selection rule: within each method and language, choose the lowest observed CER/WER and report the other metrics at that same step (Table 3). English and Chinese steps can differ and are explicitly listed; no row combines metrics from different checkpoints.
| Method | Split | Step | n | Error | Cases | SIM |
| VoxCPM2 base | EN | 0 | 1088 | 1.710 | 1 | 66.789 |
| VoxCPM2 base | ZH | 0 | 2020 | 1.006 | 1 | 72.676 |
| GROW | EN | 800 | 1088 | 1.621 | 2 | 67.588 |
| GROW | ZH | 800 | 2020 | 0.911 | 3 | 73.361 |
| LoCA w/o thresholds | EN | 400 | 1088 | 1.544 | 0 | 67.336 |
| LoCA w/o thresholds | ZH | 600 | 2020 | 0.914 | 0 | 74.140 |
| LoCA | EN | 800 | 1088 | 1.530 | 0 | 69.362 |
| LoCA | ZH | 400 | 2020 | 0.800 | 0 | 75.101 |
| LoCA w/o local | EN | 400 | 1088 | 1.572 | 0 | 67.539 |
| LoCA w/o local | ZH | 800 | 2020 | 0.858 | 2 | 73.180 |
Table 3. Best observed standard Seed results per method and language. Each row is selected by its own content error, with step shown. Error is CER for ZH and WER for EN; Cases is the severe proxy count and SIM is scaled by 100.
Severe Failure Categories and Detection
Table 4 shows Seed-hard categories at the reported checkpoints. High error means utterance CER above 50%; tail and internal silence thresholds are 1.2 and 1.5 seconds. Categories overlap, so their counts do not sum to the union. Leading silence and recognition failures alone do not trigger SFR.
| Method | Step | High err. | Silent | Tail | Internal | Cap | Union |
| VoxCPM2 base | 0 | 5 | 0 | 0 | 0 | 3 | 6 |
| GROW | 400 | 8 | 0 | 0 | 0 | 7 | 9 |
| LoCA w/o thresholds | 600 | 5 | 0 | 0 | 0 | 3 | 6 |
| LoCA | 1200 | 5 | 0 | 0 | 0 | 1 | 5 |
| LoCA w/o local | 1000 | 8 | 0 | 0 | 1 | 4 | 10 |
Table 4. Seed-hard severe-failure categories at each method’s reported checkpoint, . Union counts each flagged utterance once.
Silence detection uses mono audio, a 50-ms RMS window and a 20-ms hop. The activity threshold is 15% of the waveform’s 95th-percentile frame RMS , clipped to . No active frames means all-silent output; tail and internal durations are measured from active-frame positions. A cap hit denotes maximum-length stopping. These flags are reproducible diagnostics: a cap hit need not imply omitted content, and ASR differences can include phonetic or formatting ambiguities.
Multilingual Generalization and Similarity
Figure 3 reports all three metrics for the full CV3 selection and Russian. Russian was excluded from RL training, while the pretrained backbone can already synthesize it. At the reported checkpoints, Russian WER is 6.219% for the base, 5.904% for GROW and 4.723% for LoCA. The improvements therefore reflect post-training transfer rather than first acquisition of the language.
Figure 3. Full-subset CV3 and Russian trajectories, including speaker similarity. Russian is outside the six-language RL training pool. All plotted points are reported evaluations; the base is a horizontal reference.
The radar chart divides each split’s base error by the method’s corresponding error (Figure 4). This avoids directly mixing language-dependent CER/WER scales; a radius above one indicates improvement. The chart is a profile overview, not an aggregate score or a claim based on polygon area. Exact errors and checkpoint steps are retained in Table 6.
Figure 4. Inverse error ratio (base error / method error), at each system’s reported checkpoint. Larger is better; 1 is the base. Asterisks mark Spanish, Italian and Russian, excluded from the RL training pool. Hard Chinese/English are separate tasks, not extra languages.
| Method | Step | zh | en | de | es | fr | it | ru | h-zh | h-en |
| VoxCPM2 base | 0 | 3.996 | 5.193 | 5.551 | 4.571 | 10.657 | 5.023 | 6.219 | 9.952 | 7.665 |
| GROW | 400 | 3.398 | 4.997 | 4.846 | 4.491 | 10.314 | 4.275 | 5.904 | 10.402 | 6.053 |
| LoCA w/o thresholds | 600 | 3.396 | 5.017 | 4.500 | 5.306 | 10.638 | 3.917 | 5.300 | 8.501 | 5.333 |
| LoCA | 1200 | 3.238 | 4.507 | 3.878 | 3.608 | 9.027 | 3.731 | 4.723 | 7.186 | 7.143 |
| LoCA w/o local | 1000 | 3.402 | 4.806 | 4.562 | 4.284 | 10.502 | 4.803 | 5.617 | 8.120 | 7.593 |
Table 6. Exact CV3 language errors at each method’s selected checkpoint. Chinese uses CER; other splits use WER. All values are percentages; h-zh/h-en are the hard subsets.
| Method | Step | CV3 SIM | CV3 P808 | Seed-h SIM | Seed-h P808 |
| VoxCPM2 base | 0 | 67.168 | 3.570 | 67.632 | 3.691 |
| GROW | 400 | 67.562 | 3.580 | 67.346 | 3.664 |
| GROW | 800 | 67.558 | 3.556 | 67.135 | 3.626 |
| GROW | 1200 | 66.941 | 3.537 | 67.559 | 3.680 |
| GROW | 1600 | 66.363 | 3.551 | 65.767 | 3.694 |
| GROW | 2000 | 66.662 | 3.581 | 66.854 | 3.640 |
| LoCA w/o thresholds | 600 | 67.722 | 3.605 | 69.179 | 3.753 |
| LoCA | 800 | 68.676 | 3.601 | 68.046 | 3.666 |
| LoCA | 1200 | 69.137 | 3.636 | 68.951 | 3.641 |
| LoCA | 1600 | 68.724 | 3.619 | 68.853 | 3.649 |
| LoCA | 2000 | 68.144 | 3.604 | 67.868 | 3.698 |
| LoCA w/o local | 1000 | 67.525 | 3.564 | 67.151 | 3.648 |
Table 5. Auxiliary quality metrics at the selected checkpoints and representative training steps for GROW and LoCA. SIM and P808 are higher-is-better; SIM backends differ between CV3 and Seed.
References
- [1]Philip Anastassiou, Jiawei Chen, Jitong Chen, Yuanzhe Chen, Zhuo Chen, Ziyi Chen, Jian Cong, Lelai Deng, Chuang Ding, Lu Gao, Mingqing Gong, Peisong Huang, Qingqing Huang, Zhiying Huang, Yuanyuan Huo, Dongya Jia, Chumin Li, Feiya Li, Hui Li, and 27 others. 2024. Seed-tts: A family of high-quality versatile speech generation models. Preprint, arXiv:2406.02430.arxiv.org/abs/2406.02430
- [2]ByteDance Speech. 2024. Seed-TTS-Eval. Software and evaluation data. Accessed October 9, 2026.
- [3]Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Xiangzhang Yu, and Furu Wei. 2022. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. Preprint, arXiv:2110.13900.DOI
- [4]Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Qian Chen, and Jiajun Qi. 2023. An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification. In Interspeech 2023, pages 2228–2232.arxiv.org/abs/2305.12838
- [5]Jiayan Cui, Zhihan Yang, Naihan Li, Jiankun Tian, Xingyu Ma, Yi Zhang, Guangyu Chen, Runxuan Yang, Yuqing Cheng, Yizhi Zhou, Guochen Yu, Xiaotao Gu, and Jie Tang. 2025. GLM-TTS technical report. Preprint, arXiv:2512.14291.arxiv.org/abs/2512.14291
- [6]Zheming Fu, Ruizhe He, Wei Shang, Xiaoxiao Ma, Lei Wang, Chang Liu, and Siming Fu. 2026. FlowAWR: Online adaptive flow reinforcement via advantage-weighted rectification. Preprint, arXiv:2606.30376.arxiv.org/abs/2606.30376
- [7]Changfeng Gao, Zhihao Du, and Shiliang Zhang. 2025. Differentiable reward optimization for llm based tts system. Preprint, arXiv:2507.05911.arxiv.org/abs/2507.05911
- [8]Zhifu Gao, ShiLiang Zhang, Ian McLoughlin, and Zhijie Yan. 2022. Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition. In Interspeech 2022, pages 2063–2067.arxiv.org/abs/2206.08317
- [9]Haorui He, Zengqiang Shang, Chaoren Wang, Xuyuan Li, Yicheng Gu, Hua Hua, Liwei Liu, Chen Yang, Jiaqi Li, Peiyang Shi, Yuancheng Wang, Kai Chen, Pengyuan Zhang, and Zhizheng Wu. 2024. Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation. Preprint, arXiv:2407.05361.arxiv.org/abs/2407.05361
- [10]Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. Preprint, arXiv:2207.12598.arxiv.org/abs/2207.12598
- [11]Dongya Jia, Zhuo Chen, Jiawei Chen, Chepeng Du, Jian Wu, Jian Cong, Xiaobin Zhuang, Chumin Li, Zhen Wei, Yuping Wang, and Yuxuan Wang. 2025. Ditar: Diffusion transformer autoregressive modeling for speech generation. Preprint, arXiv:2502.03930.arxiv.org/abs/2502.03930
- [12]Rikuto Kotoge and Yuichi Sasaki. 2026. Data-efficient targeted token-level preference optimization for LLM-based text-to-speech. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 719–726, San Diego, California, United States. Association for Computational Linguistics.
- [13]Branislav Kveton, Anup Rao, Subhojyoti Mukherjee, Krishna Kumar Singh, and Viet Dac Lai. 2026. AdvantageFlow: Regularized advantage-weighted rl in flow models. Preprint, arXiv:2605.26013v2.arxiv.org/abs/2605.26013
- [14]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2023. Flow matching for generative modeling. Preprint, arXiv:2210.02747.arxiv.org/abs/2210.02747
- [15]Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. Preprint, arXiv:1711.05101.arxiv.org/abs/1711.05101
- [16]OpenBMB. 2026. VoxCPM2: Official 2b-parameter model release. Model weights and model card. Accessed October 13, 2026.
- [17]QwenAudio. 2026. CV3-Eval: Evaluation tools for text-to-speech. Software repository. Accessed October 9, 2026.
- [18]Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 28492–28518. PMLR.arxiv.org/abs/2212.04356
- [19]Chandan K. A. Reddy, Vishak Gopal, and Ross Cutler. 2021. DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors. In ICASSP 2021 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 6493–6497.arxiv.org/abs/2010.15258
- [20]Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, Jin Xu, Jingren Zhou, and Junyang Lin. 2026. Qwen3-asr technical report. Preprint, arXiv:2601.21337.arxiv.org/abs/2601.21337
- [21]Guanrou Yang, Tian Tan, Qian Chen, Ziyang Ma, Yakun Song, Zhikang Niu, Qi Chen, Wenming Tu, Haitao Li, Shan Yang, and Xie Chen. 2026. GROW: Group-relative advantage-weighted on-policy reinforcement learning of autoregressive-diffusion text-to-speech model. Preprint, arXiv:2608.03215.arxiv.org/abs/2608.03215
- [22]Zai Org. 2025. GLM-TTS: Public training implementation. Software, commit 4b944f4be7b6c55454751715081f4dc83992897a.
- [23]Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Renjie Yu, Jiancheng Gui, Jiahe Wang, Ziyang Wang, Xudong Shen, Runchuan Ye, Zhisheng Zhang, Jiyuang Zhou, Bingsong Bai, Weiyue Sun, Mengyuan Deng, Qundong Shi, Zhiyong Wu, and Zhiyuan Liu. 2026. Voxcpm2 technical report. Preprint, arXiv:2606.06928.arxiv.org/abs/2606.06928
- [24]Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Renjie Yu, Ziyang Wang, Runchuan Ye, Weiyue Sun, Jiancheng Gui, Kehan Li, Zhiyong Wu, and Zhiyuan Liu. 2025. VoxCPM: Tokenizer-free speech generation and true-to-life voice cloning. Preprint, arXiv:2509.24650.arxiv.org/abs/2509.24650



