Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Synthetic data that steadies low-resource spoken language models also flattens their prosody once the synthetic share grows large; self-alignment can restore both accuracy and expressivity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 23:21 UTC pith:AFR62CZK

load-bearing objection Solid empirical paper: synthetic scaling really does trade stability for expressivity in low-resource SLMs, and the two self-alignment fixes work on Thai/Lao. the 3 major comments →

arxiv 2605.27383 v1 pith:AFR62CZK submitted 2026-04-10 cs.CL cs.AI

Bridging the Stability-Expressivity Gap: Synthetic Data Scaling and Preference Alignment for Low-Resource Spoken Language Models

Yizhong Geng , Yanliang Li , Jinghan Yang , Tianhan Jiang , Boxun An , Ya Li , Xiaoyu Shen This is my paper
classification cs.CL cs.AI
keywords spoken language modelssynthetic datalow-resource TTSStability-Expressivity Gapprosodyself-alignmentzero-shot voice cloningThai and Lao
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Spoken Language Models generate speech from discrete tokens without hand-built pronunciation rules, but they need large amounts of transcribed audio. For languages with little real speech, the usual fix is to train on synthetic speech from existing TTS engines. This paper shows that the fix is two-edged: as the synthetic share rises, word-error rate falls steadily, yet prosodic diversity, naturalness and speaker similarity peak near a fifty-percent mix and then collapse—a non-monotonic trade-off the authors name the Stability-Expressivity Gap, or Synthetic Erosion. They close the gap with two annotation-free self-alignment loops. Disentanglement-Guided Self-Alignment exploits the model’s built-in separation of speaking style from voice identity to build its own preference pairs; Temperature-Driven Self-Critique explores many decoding temperatures and keeps only the outputs an automatic recognizer accepts. On Thai the resulting system beats commercial engines; on Lao it produces the first zero-shot voice-cloning TTS. The result matters because it turns synthetic data from a blunt instrument into a controllable path to high-fidelity speech technology for languages that lack large native corpora.

Core claim

When Spoken Language Models for low-resource languages are scaled with synthetic speech, phonetic accuracy improves monotonically with the synthetic ratio, yet token-level prosodic entropy, naturalness MOS and speaker similarity peak near a 50 percent synthetic mix and then collapse—Synthetic Erosion. Two self-alignment methods that generate their own preference pairs without human labels, Disentanglement-Guided Self-Alignment and Temperature-Driven Self-Critique, recover expressivity while preserving the phonetic gains, yielding systems that surpass commercial baselines on Thai and enable the first zero-shot voice cloning for Lao.

What carries the argument

The Stability-Expressivity Gap (Synthetic Erosion), diagnosed by token entropy of the autoregressive speech tokens; Disentanglement-Guided Self-Alignment (DGSA), which builds preference pairs by toggling style tokens while holding timbre fixed; and Temperature-Driven Self-Critique (TDSC), which generates candidates across temperatures and filters them with ASR-based criteria for iterative self-refinement.

Load-bearing premise

The central claim rests on treating the entropy of discrete speech tokens as a reliable proxy for how expressive the speech sounds to listeners, and on automatic speech recognizers being accurate enough both to filter synthetic data and to judge the model’s own outputs.

What would settle it

Train the same backbone with fixed real data while sweeping the synthetic ratio from near zero to one hundred percent; if prosodic entropy, naturalness MOS and speaker MOS keep rising or stay flat past the fifty-percent point instead of peaking and collapsing, Synthetic Erosion does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic ratio should be treated as a tunable hyperparameter, not “more is always better,” with corrective alignment applied once the critical mix is exceeded.
  • Flow-Matching architectures that already separate prosody from timbre can construct preference pairs for free by re-synthesizing the same content with and without style tokens.
  • Any language that has even a usable ASR system can bootstrap high-quality TTS and zero-shot cloning from purely synthetic data via temperature-driven self-critique.
  • Commercial-grade synthesis and voice cloning become feasible for Thai and Lao without collecting massive new native speech corpora.
  • The same non-monotonic erosion pattern is expected whenever low-entropy teacher speech dominates training of generative speech models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The mixture-entropy argument may apply to other generative domains (music, video) that mix real data with deterministic synthetic teachers.
  • If token entropy reliably tracks perceived prosody, it can serve as an online training monitor that automatically triggers expressivity recovery.
  • Replacing the ASR judge in TDSC with unsupervised or cross-lingual recognition would extend the pure-synthetic pipeline to languages that still lack any baseline recognizer.
  • Dual-objective preference alignment that isolates distinct failure modes may generalize to any generative architecture with factorized latent controls.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper characterizes a Stability-Expressivity Gap (Synthetic Erosion) that arises when scaling Spoken Language Models for low-resource languages with synthetic data: phonetic stability (WER) improves monotonically with synthetic ratio α while prosodic metrics (token entropy Hp, NMOS, SMOS) peak near α≈50% and then collapse. It proposes two self-alignment methods that exploit Flow-Matching SLM architecture—Disentanglement-Guided Self-Alignment (DGSA) for languages with limited real speech (Thai) via prosody-timbre style toggling and dual DPO objectives, and Temperature-Driven Self-Critique (TDSC) for extreme scarcity (Lao) via multi-temperature candidate generation, ASR filtering, and iterative SFT+DPO. Experiments on Thai (mixed real+synthetic) and Lao (pure synthetic) report recovery of both axes, outperformance of open-source and commercial baselines (ElevenLabs, Gemini, Azure) on standard TTS and zero-shot cloning, and the first Lao zero-shot voice cloning system.

Significance. If the empirical pattern and recovery methods hold, the work supplies a practical, annotation-light pathway for high-fidelity SLM synthesis in the long tail of languages, where synthetic data is already the default. The systematic multi-metric scaling study (Table 1, Figs. 1/4), controlled validation that Hp tracks acoustic and perceptual prosody under matched WER (Table 2), clean stage separation of SFT from preference alignment, dual-objective ablations, TDSC iteration dynamics, and statistical reporting (95% CIs, paired t-tests) are concrete strengths. Enabling zero-shot Lao cloning and competitive or superior commercial results on Thai further raise the applied value for linguistic inclusion. The mixture-entropy argument is only qualitative intuition and is not required for the claims.

major comments (3)
  1. [§5.4 Table 4 vs §5.5 Table 5 / Table 6] Section 5.4 Table 4 reports TDSC Lao NMOS = 3.94 ± 0.07 (and the surrounding text states a rise from 3.1 to 3.9), while Section 5.5 Table 5 and the accompanying prose claim NMOS = 4.53 ± 0.06 and superiority over Gemini Flash (4.12). Table 6 reverts to 3.94. These mutually inconsistent numbers for the identical system and metric directly underwrite the SOTA and commercial-outperformance claims; they must be reconciled with a single evaluation protocol before the central results can be trusted.
  2. [§4 / §5.4 / Limitations] The load-bearing diagnostic and filtering role of ASR (Whisper-large-v3 for Thai; Dolphin-small at the paper’s own 21.5% WER for Lao) is acknowledged in Limitations but under-analyzed for TDSC. Because TDSC constructs both the accepted set G and the preference pairs (yw, yl) solely from this judge (Eq. 16 and surrounding text), residual ASR error can systematically bias the preference signal toward ASR-friendly rather than human-preferred prosody. A sensitivity study (e.g., injecting controlled ASR noise or reporting inter-ASR agreement) is needed to confirm that the reported gains are not artifacts of the judge.
  3. [§2 / §3 Eqs. (13)–(14) / §5.2] Hp is introduced as an architecture-justified proxy for prosodic diversity (Section 2) and is validated under matched-WER conditions (Table 2). However, the dynamic weight schedule of DGSA (Eqs. 13–14) and the temperature curriculum of TDSC both treat the location of the Hp peak (α*≈50%) as a transferable constant. Because α* is read off the same Thai scaling curve that is later used for evaluation, a modest leave-one-ratio-out or cross-language check would strengthen the claim that the schedule is not over-fit to the demonstration data.
minor comments (5)
  1. [Figure 1 / Figure 4] Figure 1 and Figure 4 are informative but the axis labels and legend fonts become hard to read at print scale; consider enlarging or splitting the multi-metric panel.
  2. [§4 / Appendix E] The definition of repetition rate (k=4 consecutive identical tokens) appears only in Appendix E; a one-sentence pointer in the main text near Eq. (16) would help readers.
  3. [§5.1 / Appendix D] Commercial API versions and the exact freeze date (25 Jan 2025) are given in Appendix D; a short footnote in Section 5.1 would improve reproducibility without forcing readers into the appendix.
  4. [Throughout] Minor typographic inconsistencies appear (e.g., “CosyV oice” vs “CosyVoice”, occasional missing spaces after citations). A final proof-reading pass would clean them.
  5. [§5.2] The pure-synthetic Lao baseline (Table 1 α=100% and Table 4 SFT) is useful cross-language corroboration of Synthetic Erosion; a single sentence in Section 5.2 explicitly linking the two would tighten the narrative.

Circularity Check

0 steps flagged

No significant circularity: empirical scaling curves and self-alignment methods are self-contained against external benchmarks and independent acoustic/human metrics.

full rationale

The paper's central claims (non-monotonic Stability-Expressivity Gap under synthetic ratio α, Synthetic Erosion beyond α≈50%, and recovery via DGSA/TDSC) rest on direct experimental measurements (Table 1, Figures 1/4, controlled Hp study in Table 2, ablations in Tables 3/7, TDSC dynamics in Figure 6/Table 4) rather than any closed-form derivation that reduces to its own inputs. The mixture-entropy argument (Appendix B, Lemmas B.1–B.2) is classical Shannon entropy concavity applied only as qualitative intuition for the observed two-phase pattern; it is explicitly non-load-bearing and does not force the empirical α* or the methods. Hp is introduced as a diagnostic proxy justified by the external CosyVoice/Vevo architectural separation of AR tokens (prosody) from Flow-Matching (timbre), then independently validated against F0 statistics, energy variation, and human MOS on matched-WER pairs (Table 2); it is not defined circularly from the target MOS. DGSA constructs preference pairs by toggling the style token while freezing timbre embeddings (Eqs. 6–9), then applies standard DPO; the pairs are generated from a frozen SFT checkpoint and never re-enter the SFT objective. TDSC uses an external ASR judge (Whisper/Dolphin) plus length/repetition filters to accept/reject multi-temperature candidates before SFT+DPO; this is iterative self-improvement with an external filter, not pure self-reward or a fitted constant renamed as prediction. α*=0.5 is an observed empirical heuristic from the scaling curve, not a uniqueness theorem or self-citation chain. No self-definitional equations, no fitted-input-called-prediction, and no load-bearing uniqueness imported from overlapping authors appear. Mild self-improvement loop risk exists in TDSC but does not reduce the reported Thai/Lao gains (vs. ElevenLabs/Gemini) to construction. Score 1 reflects only the ordinary use of the demonstration curve to set the scheduling threshold.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 4 invented entities

The central claims rest on architectural assumptions imported from CosyVoice/Vevo, classical mixture-entropy concavity, and several hand-chosen thresholds (α*, temperature set, WER/repetition filters) that are not derived from first principles. No new physical entities are postulated; the named phenomena (Synthetic Erosion, Stability-Expressivity Gap) are descriptive labels for observed scaling behavior. Free parameters are the usual ML hyper-parameters plus the critical synthetic ratio used for dynamic weighting.

free parameters (4)
  • critical synthetic ratio α* = 0.5
    Set to 0.5 from the observed peak of Hp/NMOS on the Thai scaling curve; used both to interpret Synthetic Erosion and to schedule DGSA λe/λs. Not held out.
  • TDSC temperature set and curriculum = {0.7, 1.0, 0.8+0.1k}
    T∈{0.7,1.0,Tmax} with Tmax=0.8+0.1k chosen by hand; controls the stability-expressivity exploration that TDSC claims to balance.
  • ASR filtering thresholds (τw, τr, γmin, γmax) = τw=40%/50%, τr=10%, γ=[0.5,2.0]
    WER <40%/50%, repetition <10%, length ratio [0.5,2.0]×|x| define which self-generated samples enter TDSC training; directly determine the accepted set G.
  • DPO β and DGSA λe/λs schedule = β=0.1; λe linear in α above 0.5
    β=0.1 and the linear crossover λe=max(0,(α-α*)/(1-α*)) are design choices that control preference strength and when expressivity correction activates.
axioms (4)
  • domain assumption In Flow-Matching SLMs the autoregressive tokens primarily encode content+prosody while the flow-matching decoder controls timbre via independent embeddings (prosody-timbre separation).
    Imported from CosyVoice 2 / Vevo citations; underpins both Hp as prosodic proxy and the dual-mode generation used by DGSA (Section 2–3).
  • standard math Shannon entropy of a mixture distribution is strictly concave when the component distributions differ, implying a unique maximizer α*.
    Classical result restated in Appendix B to give qualitative intuition for the two-phase scaling curve; not claimed as a novel theorem.
  • domain assumption A usable (even moderate-accuracy) ASR system exists for the target language and can serve as both data filter and TDSC judge.
    Explicitly stated in Limitations; TDSC and all WER numbers collapse without it. Dolphin-small at 21.5% WER on Lao is accepted as sufficient.
  • domain assumption External deterministic TTS engines produce lower-entropy token distributions than human speech (H(psyn)<H(preal)).
    Stated in Section 2 as the driver of Synthetic Erosion; supported by the observed scaling but treated as given for the mixture analysis.
invented entities (4)
  • Stability-Expressivity Gap / Synthetic Erosion independent evidence
    purpose: Name the observed non-monotonic trade-off when synthetic data ratio exceeds a critical value.
    Descriptive label for empirical scaling behavior, not a new physical mechanism. Independent evidence is the multi-metric curves themselves.
  • Disentanglement-Guided Self-Alignment (DGSA) independent evidence
    purpose: Self-construct preference pairs by toggling style tokens while holding timbre fixed, then apply dual DPO.
    New training procedure; falsifiable by ablation (Table 3/7) and by whether identity-consistent pairs outperform random pairing.
  • Temperature-Driven Self-Critique (TDSC) independent evidence
    purpose: Closed-loop multi-temperature generation + ASR filtering + SFT/DPO for pure-synthetic regimes.
    New procedure; falsifiable by iteration dynamics (Fig. 6) and component ablations (Table 8).
  • Prosodic Entropy Hp independent evidence
    purpose: Lightweight automatic proxy for prosodic diversity of generated speech tokens.
    Defined as ordinary token entropy; claimed correlation with NMOS is empirical (Fig. 1, Table 2), not definitional.

pith-pipeline@v1.1.0-grok45 · 28551 in / 3820 out tokens · 40847 ms · 2026-07-12T23:21:14.213028+00:00 · methodology

0 comments
read the original abstract

Spoken Language Models (SLMs) have emerged as a promising paradigm for speech synthesis by bypassing explicit grapheme-to-phoneme pipelines. However, their effectiveness in low-resource languages remains fundamentally limited by the scarcity of transcribed speech. In practice, synthetic data has become the primary strategy for scaling SLMs in such settings, providing reliable phonetic supervision when real data is insufficient. In this work, we show that this reliance introduces a fundamental trade-off, which we term the Stability-Expressivity Gap: while synthetic data improves phonetic accuracy, it progressively suppresses prosodic variability, ultimately leading to a collapse of expressivity (Synthetic Erosion). To bridge this gap, we propose two self-alignment frameworks. Disentanglement-Guided Self-Alignment (DGSA) recovers expressivity for complex languages by exploiting prosody-timbre separation. For regimes where authentic references are exceptionally limited, Temperature-Driven Self-Critique (TDSC) stabilizes generation through automated exploration and filtering. Our approach outperforms strong commercial systems, including ElevenLabs and Gemini Pro, and enables the first zero-shot voice cloning capability for Lao.

Figures

Figures reproduced from arXiv: 2605.27383 by Boxun An, Jinghan Yang, Tianhan Jiang, Xiaoyu Shen, Ya Li, Yanliang Li, Yizhong Geng.

Figure 1
Figure 1. Figure 1: Scaling behavior of objective metrics as synthetic data ratio α increases. WER decreases monotonically, indicating improved stability. In contrast, token entropy Hp, repetition rate, NMOS, and SMOS exhibit non-monotonic trends—peaking around α ≈ 50% before degrading, revealing the Synthetic Erosion phenomenon. Notably, Hp tracks NMOS, supporting its use as a lightweight proxy for prosodic diversity. The tr… view at source ↗
Figure 2
Figure 2. Figure 2: Disentanglement-Guided Self-Alignment (DGSA). Flow-Matching SLMs separate prosody (Text-Speech LM) from timbre (Flow-Matching Transformer). Enabling the style token produces expressive output y expr; disabling it yields stable but flat output y stab . DGSA aligns both toward real speech y real via dual preference objectives. why naive scaling fails and motivating the alignment-based corrections we propose … view at source ↗
Figure 3
Figure 3. Figure 3: Temperature-Driven Self-Critique (TDSC) For each input, the model generates candidates at temperatures T ∈ {low, mid, high}, spanning conservative (stable) to exploratory (expressive) outputs. The Judge Model filters candidates by WER, length, and repetition criteria, yielding accepted (G) and rejected (R) sets for preference-based refinement. preference triplets starting from the same SFT checkpoint. This… view at source ↗
Figure 4
Figure 4. Figure 4: Stability-Expressivity trade-off space. Each point repre￾sents a model trained with different synthetic data ratios. Lower WER (rightward) indicates better stability; higher Hp (upward) indicates better expressivity. The 300h configuration achieves the best balance, while excessive synthetic data (1200h, 1500h) sacri￾fices expressivity for marginal stability gains. 5.2. Scaling Experiments We first examine… view at source ↗
Figure 5
Figure 5. Figure 5: Dynamic weight scheduling and Hp recovery. Below α ∗ = 50%, λe = 0 (no correction needed). Beyond α ∗ , λe activates and ∆Hp scales proportionally [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: TDSC iteration dynamics over 5 refinement rounds. WER decreases steadily while Hp rises in later iterations as Tmax expands. Pass rate increases from 23% to 62%, indicating progres￾sive quality improvement. 5.4. TDSC Evaluation We evaluate TDSC on Lao, a low-resource language with exceptionally limited real speech corpus. The model is trained entirely on 1,500h of synthetic data generated via cross-lingual… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 12 linked inside Pith

  1. [1]

    C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al

    Barrault, L., Chung, Y .-A., Meglioli, M. C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al. Seamlessm4t: Massively multilin- gual & multimodal machine translation.arXiv preprint arXiv:2308.11596,

    Pith/arXiv arXiv

  2. [2]

    Tts- transducer: End-to-end speech synthesis with neural transducer

    Bataev, V ., Ghosh, S., Lavrukhin, V ., and Li, J. Tts- transducer: End-to-end speech synthesis with neural transducer. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

    2025

  3. [3]

    Scaling low- resource mt via synthetic data generation with llms.arXiv preprint arXiv:2505.14423,

    de Gibert, O., Attieh, J., Vahtola, T., Aulamo, M., Li, Z., V´azquez, R., Hu, T., and Tiedemann, J. Scaling low- resource mt via synthetic data generation with llms.arXiv preprint arXiv:2505.14423,

    arXiv

  4. [4]

    and Yu, K

    Du, C. and Yu, K. Speaker augmentation for low resource speech recognition. InICASSP 2020-2020 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 7719–7723. IEEE,

    2020

  5. [5]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models.arXiv preprint arXiv:2412.10117,

    Du, Z., Wang, Y ., Chen, Q., Shi, X., Lv, X., Zhao, T., Gao, Z., Yang, Y ., Gao, C., Wang, H., et al. Cosyvoice 2: Scalable streaming speech synthesis with large language models.arXiv preprint arXiv:2412.10117,

    Pith/arXiv arXiv

  6. [6]

    Gao, X., Zhang, C., Chen, Y ., Zhang, H., and Chen, N. F. Emo-dpo: Controllable emotional speech synthe- sis through direct preference optimization. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

    2025

  7. [7]

    Surveying the technology support of languages

    Gr¨utzner-Zahn, A., Gaspari, F., Giagkou, M., Hegele, S., Way, A., and Rehm, G. Surveying the technology support of languages. InProceedings of the Second International Workshop Towards Digital Language Equality (TDLE): Focusing on Sustainability@ LREC-COLING 2024, pp. 1–17,

    2024

  8. [8]

    Univoice: Unifying autoregres- sive asr and flow-matching based tts with large language models.arXiv preprint arXiv:2510.04593,

    Guan, W., Niu, Z., Jiang, Z., Wang, K., Chen, P., Hong, Q., Li, L., and Chen, X. Univoice: Unifying autoregres- sive asr and flow-matching based tts with large language models.arXiv preprint arXiv:2510.04593,

    Pith/arXiv arXiv

  9. [9]

    Low-resource expressive text-to- speech using data augmentation

    10 Bridging the Stability-Expressivity Gap: Synthetic Data Scaling and Preference Alignment for Low-Resource SLMs Huybrechts, G., Merritt, T., Comini, G., Perz, B., Shah, R., and Lorenzo-Trueba, J. Low-resource expressive text-to- speech using data augmentation. InICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (I...

    2021

  10. [10]

    J., Cao, Y ., Chiu, C.-C., Ari, N., Laurenzo, S., and Wu, Y

    Jia, Y ., Johnson, M., Macherey, W., Weiss, R. J., Cao, Y ., Chiu, C.-C., Ari, N., Laurenzo, S., and Wu, Y . Leveraging weakly supervised data to improve end-to-end speech- to-text translation. InICASSP 2019-2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 7180–7184. IEEE,

    2019

  11. [11]

    Naturalspeech 3: Zero-shot speech synthesis with factorized codec and dif- fusion models.arXiv preprint arXiv:2403.03100,

    Ju, Z., Wang, Y ., Shen, K., Tan, X., Xin, D., Yang, D., Liu, Y ., Leng, Y ., Song, K., Tang, S., et al. Naturalspeech 3: Zero-shot speech synthesis with factorized codec and dif- fusion models.arXiv preprint arXiv:2403.03100,

    Pith/arXiv arXiv

  12. [12]

    Self-training for end- to-end speech recognition

    Kahn, J., Lee, A., and Hannun, A. Self-training for end- to-end speech recognition. InICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), pp. 7084–7088. IEEE,

    2020

  13. [13]

    Parameter-efficient fine-tuning for low-resource text-to-speech via cross- lingual continual learning

    Kwon, K.-J., So, J.-H., and Lee, S.-H. Parameter-efficient fine-tuning for low-resource text-to-speech via cross- lingual continual learning. InProc. Interspeech, volume 2025, pp. 1613–1617,

    2025

  14. [14]

    Disco-speech: Controllable zero-shot speech generation with a disentangled speech codec.arXiv preprint arXiv:2512.13251,

    Li, T., Ge, W., Wang, Z., Cui, Z., Ma, Y ., Gao, Y ., Deng, C., Zhang, S., and Feng, J. Disco-speech: Controllable zero-shot speech generation with a disentangled speech codec.arXiv preprint arXiv:2512.13251,

    arXiv

  15. [15]

    Lao-english code-switched speech synthesis via neural codec language modeling

    Liu, Y ., Wang, L., Gao, S., Yu, Z., Dong, L., and Tian, T. Lao-english code-switched speech synthesis via neural codec language modeling. InChina National Confer- ence on Chinese Computational Linguistics, pp. 105–118. Springer, 2025a. Liu, Z., Jia, D., Wang, X., Du, C., Wang, S., Chen, Z., and Li, H. Direct preference optimization for speech autoregres-...

    arXiv

  16. [16]

    Mayer, P., Lux, F., P´erez-Gonz´alez-de Martos, A., Elizarova, A., Vanderlyn, L., V¨ath, D., and Vu, N. T. Investigating stochastic methods for prosody modeling in speech syn- thesis.arXiv preprint arXiv:2507.00227,

    Pith/arXiv arXiv

  17. [17]

    and Nikolov, N

    McGiff, J. and Nikolov, N. S. Overcoming data scarcity in generative language modelling for low-resource languages: A systematic review.arXiv preprint arXiv:2505.04531,

    Pith/arXiv arXiv

  18. [18]

    Mehta, S., Tu, R., Beskow, J., Sz´ekely, ´E., and Henter, G. E. Matcha-tts: A fast tts architecture with conditional flow matching. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 11341–11345. IEEE,

    2024

  19. [19]

    Dolphin: A large- scale automatic speech recognition model for eastern languages.arXiv preprint arXiv:2503.20212,

    Meng, Y ., Li, J., Lin, G., Pu, Y ., Wang, G., Du, H., Shao, Z., Huang, Y ., Li, K., and Zhang, W.-Q. Dolphin: A large- scale automatic speech recognition model for eastern languages.arXiv preprint arXiv:2503.20212,

    Pith/arXiv arXiv

  20. [20]

    Is synthetic data truly effective for training speech language models? InProc

    Mizumoto, T., Kojima, A., Fujita, Y ., Liu, L., and Sudo, Y . Is synthetic data truly effective for training speech language models? InProc. Interspeech 2025, pp. 1808– 1812,

    2025

  21. [21]

    V oices- tar: Robust zero-shot autoregressive tts with duration con- trol and extrapolation.arXiv preprint arXiv:2505.19462,

    11 Bridging the Stability-Expressivity Gap: Synthetic Data Scaling and Preference Alignment for Low-Resource SLMs Peng, P., Li, S.-W., Mohamed, A., and Harwath, D. V oices- tar: Robust zero-shot autoregressive tts with duration con- trol and extrapolation.arXiv preprint arXiv:2505.19462,

    Pith/arXiv arXiv

  22. [22]

    Pythainlp: Thai natural language processing in python.arXiv preprint arXiv:2312.04649,

    Phatthiyaphaibun, W., Chaovavanich, K., Polpanumas, C., Suriyawongkul, A., Lowphansirikul, L., Chormai, P., Limkonchotiwat, P., Suntorntip, T., and Udomcharoen- chaikit, C. Pythainlp: Thai natural language processing in python.arXiv preprint arXiv:2312.04649,

    Pith/arXiv arXiv

  23. [23]

    Typhoon 2: A family of open text and multimodal thai large language models.arXiv preprint arXiv:2412.13702,

    Pipatanakul, K., Manakul, P., Nitarach, N., Sirichote- dumrong, W., Nonesung, S., Jaknamon, T., Pengpun, P., Taveekitworachai, P., Na-Thalang, A., Sripaisarn- mongkol, S., et al. Typhoon 2: A family of open text and multimodal thai large language models.arXiv preprint arXiv:2412.13702,

    Pith/arXiv arXiv

  24. [24]

    M., Rath, S

    Ragni, A., Knill, K. M., Rath, S. P., and Gales, M. J. Data augmentation for low resource languages. InINTER- SPEECH 2014: 15th annual conference of the inter- national speech communication association, pp. 810–

    2014

  25. [25]

    Encoding of lexical tone in self-supervised models of spoken language

    Shen, G., Watkins, M., Alishahi, A., Bisazza, A., and Chrupała, G. Encoding of lexical tone in self-supervised models of spoken language. InProceedings of the 2024 Conference of the North American Chapter of the Associ- ation for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 4250–4261,

    2024

  26. [26]

    V ., Montalan, J

    Susanto, Y ., Hulagadri, A. V ., Montalan, J. R., Ngui, J. G., Yong, X., Leong, W. Q., Rengarajan, H., Limkonchotiwat, P., Mai, Y ., and Tjhi, W. C. Sea-helm: Southeast asian holistic evaluation of language models. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 12308–12336,

    2025

  27. [27]

    Neural codec language models are zero-shot text to speech synthesizers

    Wang, C., Chen, S., Wu, Y ., Zhang, Z., Zhou, L., Liu, S., Chen, Z., Liu, Y ., Wang, H., Li, J., et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111,

    Pith/arXiv arXiv

  28. [28]

    Theoretical proof that auto-regressive language models collapse when real-world data is a finite set.arXiv preprint arXiv:2412.14872,

    Wang, L., Shi, X., Li, G., Li, J., Zhang, X., Dong, Y ., Jiao, W., and Mei, H. Theoretical proof that auto-regressive language models collapse when real-world data is a finite set.arXiv preprint arXiv:2412.14872,

    Pith/arXiv arXiv

  29. [29]

    Speechalign: Aligning speech generation to human preferences.Advances in Neural Information Processing Systems, 37:50343–50360, 2024a

    Zhang, D., Li, Z., Li, S., Zhang, X., Wang, P., Zhou, Y ., and Qiu, X. Speechalign: Aligning speech generation to human preferences.Advances in Neural Information Processing Systems, 37:50343–50360, 2024a. Zhang, J., Zhang, X., Yang, J., Wang, Y ., Fan, F., and Wu, Z. Multi-metric preference alignment for generative speech restoration.arXiv preprint arXiv...

    arXiv

  30. [30]

    and Maleki, F

    Zheng, J. and Maleki, F. Selective classifier-free guid- ance for zero-shot text-to-speech.arXiv preprint arXiv:2509.19668,

    arXiv

  31. [31]

    H., Yip, J

    Zhou, K., Zhao, S., Ma, Y ., Zhang, C., Wang, H., Ng, D., Ni, C., Nguyen, T. H., Yip, J. Q., and Ma, B. Phonetic enhanced language modeling for text-to-speech synthesis. InProc. Interspeech 2024, pp. 3440–3444, 2024a. Zhou, Z., Liu, J., Shao, J., Yue, X., Yang, C., Ouyang, W., and Qiao, Y . Beyond one-preference-fits-all alignment: Multi-objective direct ...

    2024

  32. [32]

    demonstrates that well-designed SLM architectures can decouplewhat to speak(linguistic content),how to speak(prosody, including pitch, rhythm, and emphasis), andwho speaks(speaker timbre)(Zhang et al., 2025b; Du et al., 2024). Specifically, the autoregressive transformer generates discrete tokens that encode content and prosodic style, while the Flow-Matc...

    2024

  33. [33]

    Table 11.General training hyperparameters across different stages

    This rigorous configuration ensures that the model bootstraps from high-fidelity data while maintaining sufficient diversity. Table 11.General training hyperparameters across different stages. Hyperparameter SFT DGSA TDSC Learning Rate1×10 −5 1×10 −6 1×10 −5 Total Steps / Iterations 38k 10k 5 iter. Batch Size Dynamic (max 2,000 frames/GPU) Optimizer AdamW...

    2025

  34. [34]

    Similarly, for Lao, we address its scriptio continua nature by employing the laonlp4 library for linguistic normalization and word segmentation

    engine for word segmentation prior to calculating the Levenshtein distance. Similarly, for Lao, we address its scriptio continua nature by employing the laonlp4 library for linguistic normalization and word segmentation. The ASR backbone for Lao is Dolphin-small (Meng et al., 2025), chosen for its superior performance on Southeast Asian tonal phonology co...

    2025