The reviewed record of science sign in
Pith

arxiv: 2607.06054 · v1 · pith:ASQ6MVEY · submitted 2026-07-07 · cs.SD · cs.CL

BlueMagpie-TTS: A Token-Efficient Tokenizer, Language Model, and TTS for Taiwanese-Accent Code-Switching Speech

Ho Lam Chung , Bo-Xuan Zheng , Cheng-Chieh Huang , Cheng-Han Chang , Jung-Ching Chen , Lok-Lam Ieong , Ting-Lin Hsiao , Yu-Cheng Lee
show 3 more authors
Yi-Hsin Chung Yu-Kai Guo Hung-yi Lee
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 17:50 UTCglm-5.2pith:ASQ6MVEYrecord.jsonopen to challenge →

classification cs.SD cs.CL
keywords bluemagpie-ttstaiwanesetextacousticbarbetcode-switchinglanguagemandarin
0
0 comments X

The pith

Taiwan-localized text frontend cuts TTS errors 58% in code-switching speech

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

The paper argues that the dominant failure mode in off-the-shelf TTS for Taiwanese Mandarin—wrong accent, fragmented text, and garbled Chinese-English code-switching—stems not from the acoustic rendering engine but from the text side: the tokenizer and language model frontend that plan what to say and how to say it. To test this claim, the authors build a three-part stack. First, PangolinTokenizer, a byte-level BPE tokenizer trained on Taiwan-context data, achieves the lowest token rate (0.485 tokens/character) among nine compared tokenizers while using the smallest vocabulary. Second, Barbet, a billion-parameter Traditional-Chinese language model trained on that tokenizer, serves as the text-semantic frontend and ranks first among comparable public models on a 14-task evaluation. Third, BlueMagpie-TTS attaches Barbet to the pretrained acoustic stack of VoxCPM2 through a learned bridge module, keeping the acoustic stack fixed. On a 1,000-sentence Taiwan-localized test set, this stack lowers character error rate from 11.45% to 4.81%. Crucially, the authors isolate the frontend's contribution by comparing against a baseline that fine-tunes the same acoustic stack to the same target voice: BlueMagpie-TTS stays 25.2% below that baseline, and since the two systems share the same acoustic stack and the same voice fine-tuning, differing only in the frontend and its bridge, the remaining gain is attributable to the Taiwan-localized text side. A blind listening study with ten native speakers on 500 sentences confirms the intelligibility result: 65.6% of majority votes prefer BlueMagpie-TTS.

Core claim

The central finding is that when a strong acoustic stack is reused, a substantial portion of accent and code-switching quality in TTS is determined by the text-semantic frontend—the tokenizer and language model that plan the utterance—not by the acoustic renderer. By replacing only the frontend with a Taiwan-localized tokenizer (PangolinTokenizer) and language model (Barbet), while keeping the acoustic stack fixed and controlling for voice fine-tuning, the system achieves a 25.2% further CER reduction beyond voice adaptation alone. This means the switch-point quality at Chinese-English boundaries is largely a frontend planning problem, not an acoustic rendering problem.

What carries the argument

The bridge module is the central mechanism that makes the frontend-swap architecture work. It is a lightweight component—an RMSNorm, a linear projection, and one residual SwiGLU block with zero-initialized output—that maps the hidden states of Barbet (1,536-dimensional, in its own representation space) into the input space expected by the VoxCPM2 acoustic stack. The bridge is trained in two stages: first, both pretrained parts are frozen and the bridge learns to align the two spaces via hidden-space distillation; then the full assembled model is jointly fine-tuned on target-voice speech. Because the bridge is the only component learned from scratch, the recipe is cheap to apply to new target

If this is right

  • If the frontend-dominance claim generalizes, TTS systems for other under-represented language varieties could be localized by training a tokenizer and language model on cheap text data, then bridging to an existing acoustic stack, avoiding the cost of training or retraining acoustic models from scratch.
  • The token efficiency of the frontend propagates end-to-end: shorter token sequences mean lower inference cost and less KV-cache pressure, which matters for real-time and long-form synthesis.
  • The finding that short English abbreviations inside Mandarin matrices remain the hardest case suggests that the remaining code-switching failures are a data coverage problem in the frontend, not an architectural limitation of the acoustic stack.
  • The controlled comparison design—same acoustic stack, same voice fine-tuning, differing only in frontend—provides a template for isolating frontend contributions in other modular TTS architectures.

Where Pith is reading between the lines

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

  • The paper does not separate the tokenizer's contribution from the language model's contribution to the CER reduction, since Barbet is trained on PangolinTokenizer and the two are coupled. An ablation that swaps only the tokenizer or only the language model would clarify whether token efficiency or language-model specialization drives the gain.
  • The synthesize-then-recognize evaluation uses Breeze-ASR-25, a Taiwanese-Mandarin ASR model. If this ASR was trained on data with tokenization patterns similar to PangolinTokenizer's output, it could systematically favor BlueMagpie-TTS, inflating the CER gap. The human preference study mitigates this but only compares against the zero-shot base, not the fine-tuned baseline that isolates the fronte
  • The 1,000-sentence test set is internal and locally focused by design, so the reported CER numbers test worst-case code-switching scenarios rather than average-case performance. The absolute numbers may not generalize to broader usage distributions.

Load-bearing premise

The evaluation assumes that the ASR model used to score intelligibility (Breeze-ASR-25) is an unbiased judge across all TTS systems being compared. If this ASR model has its own bias toward text patterns produced by the Taiwan-localized frontend, it could systematically favor BlueMagpie-TTS outputs over the baseline, inflating the measured CER gap. The paper acknowledges the ASR makes errors and uses the same ASR for all systems, but does not verify that the ASR's error rate,

What would settle it

The frontend-dominance claim would be weakened if, when the same Taiwan-localized tokenizer and language model are attached to a different acoustic stack (not VoxCPM2), the CER reduction fails to replicate—suggesting the gain depends on an interaction between Barbet and VoxCPM2 specifically rather than being a general frontend effect. The claim would also be challenged if an independent ASR evaluator (not Breeze-ASR-25) produced a substantially smaller CER gap between BlueMagpie-TTS and the fine-tuned baseline, indicating the measured gap partly reflects ASR bias rather than true intelligible

Figures

Figures reproduced from arXiv: 2607.06054 by Bo-Xuan Zheng, Cheng-Chieh Huang, Cheng-Han Chang, Ho Lam Chung, Hung-yi Lee, Jung-Ching Chen, Lok-Lam Ieong, Ting-Lin Hsiao, Yi-Hsin Chung, Yu-Cheng Lee, Yu-Kai Guo.

Figure 1
Figure 1. Figure 1: Section III gives the overview. Our contributions follow the same bottom-up order. • We build PangolinTokenizer, a byte-level BPE tok￾enizer for Taiwan-context text and speech transcripts, with PangolinBench, a tokenizer benchmark with ten Taiwan-focused subsets (Section IV). It reaches 0.485 tokens/character, the lowest among eight tokenizers, with the smallest vocabulary. • We train Barbet, a billion-par… view at source ↗
Figure 1
Figure 1. Figure 1: The Taiwan-localized stack, built around token efficiency. PangolinTokenizer (orange) represents Taiwan text at a low token rate. Barbet (blue), a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustrative tokenization of one code-switching sentence. A Taiwan [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Vocabulary size against overall tokens/character. Lower-left is better. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: CER on the 1,000-sentence Taiwan-localized test set. Fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Off-the-shelf TTS systems are poorly adapted to Taiwanese Mandarin. Their accent defaults to other Mandarin variants, their tokenizers over-segment common Taiwanese text, and their pronunciation degrades at code-switching boundaries where Chinese and English alternate within one utterance. These problems share one root: the text side lacks adaptation to the Taiwanese context. We address the text side from the bottom up. PangolinTokenizer, a byte-level BPE tokenizer trained on Taiwan-context data, reaches the lowest token rate (0.485 tokens/character) with the smallest vocabulary among nine tokenizers. Barbet, a billion-parameter Traditional-Chinese language model trained on PangolinTokenizer, serves as the text-semantic frontend and ranks first among comparable public models on a 14-task evaluation. BlueMagpie-TTS attaches Barbet to the pretrained acoustic stack of VoxCPM2 through a learned bridge, keeping the acoustic stack fixed. On a 1000-sentence Taiwan-localized test set, it lowers CER from 11.45% to 4.81% and WER from 14.83% to 5.36%, relative reductions of 58.0% and 63.9%. In a blind listening study on 500 of these sentences with ten listeners, 65.6% of majority votes prefer BlueMagpie-TTS.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 7 minor

Summary. The paper presents BlueMagpie-TTS, a Taiwanese-accent code-switching TTS system built from three components: (1) PangolinTokenizer, a byte-level BPE tokenizer trained on Taiwan-context data; (2) Barbet, a 1.09B-parameter Traditional-Chinese language model serving as the text-semantic frontend; and (3) a reused VoxCPM2 acoustic stack connected via a learned bridge module. The central empirical claim is that replacing the text-semantic frontend with the Taiwan-localized stack, while keeping the acoustic stack fixed, lowers CER from 11.45% (zero-shot base) to 4.81%, and remains 25.2% below a fine-tuned baseline (6.43%) that shares the same acoustic stack and target-voice fine-tuning. The paper also reports tokenizer efficiency across eight baselines on a new benchmark (PangolinBench), standalone LM quality on TAIDE-14, and a human preference study showing 65.6% majority-vote preference for BlueMagpie-TTS over the zero-shot base.

Significance. The paper addresses a real gap: off-the-shelf TTS systems are poorly adapted to Taiwanese Mandarin and intra-sentential Chinese-English code-switching. The controlled comparison design—zero-shot base, fine-tuned base, and BlueMagpie-TTS sharing the same acoustic stack and fine-tuning—is well constructed and isolates the frontend effect more cleanly than a simple before-after comparison. The tokenizer benchmark with explicit pass criteria and the deterministic streaming BPE trainer are practical contributions. The human preference study with ten listeners and a sign test provides convergent evidence alongside the ASR-based metric. The recipe of reusing a pretrained acoustic stack and training only a bridge plus short fine-tuning is cost-effective and reproducible. The main concern is whether the load-bearing 25.2% claim is fully validated given the evaluation methodology.

major comments (2)
  1. §VII-A, §VII-C: The 25.2% frontend-isolated CER reduction (6.43% → 4.81%) rests on the synthesize-then-recognize loop using Breeze-ASR-25, which is itself a Taiwan-localized model. The paper states that using the same ASR for all systems keeps the comparison fair, but same-ASR does not guarantee unbiased comparison if the ASR has differential accuracy across acoustic regimes. If BlueMagpie-TTS produces speech whose accent, prosody, or code-switching timing more closely matches Breeze-ASR-25's training distribution, the ASR may transcribe BlueMagpie-TTS outputs more accurately than baseline outputs even if the actual intelligibility difference is smaller. The human preference study (§VII-D) partially addresses evaluator bias, but it only compares BlueMagpie-TTS against the zero-shot base (11.45% CER), not against the fine-tuned baseline (6.43% CER) that isolates the frontend effect. Thus,
  2. §VII-B: The 1,000-sentence test set is internal and not publicly available. The paper acknowledges this ('so the numbers support relative comparison'), but the central CER claim would be substantially strengthened by either releasing the test set, reporting results on an external benchmark, or at minimum providing summary statistics (sentence length distribution, code-switching rate, English span types) that allow readers to assess generalizability. Without this, it is unclear whether the 25.2% gap would hold on other Taiwan-localized test sets. This is load-bearing because the headline result is specific to this test set.
minor comments (7)
  1. Abstract (standalone): says 'eight tokenizers' in one place and 'nine tokenizers' in another. The body and Table I show eight baselines plus PangolinTokenizer (nine total). The abstract should say 'eight' consistently when referring to baselines, or 'nine' when including PangolinTokenizer.
  2. §IV-E, Eq. (1): The vocabulary efficiency score is described as 'an empirical formula... not a derived quantity.' The formula (1/(tokens per character)) / (vocabulary size / 100000) is dimensionally a rate scaled by an arbitrary constant. A brief justification for the 100000 scaling or a reference to prior work using this metric would help readers interpret the scores in Table II.
  3. Table I: The caption says 'BEST IN BOLD, SECOND-BEST UNDERLINED' but the formatting of bold and underline is not visible in the text rendering. Please verify these are correctly marked in the final format.
  4. §V-B, Table V: The model has 28 decoder layers with a 4-layer cycle pattern (global, sliding-window, Mamba). 28 is not divisible by 4. Please clarify how the pattern wraps or whether the last cycle is truncated.
  5. §VI-A: The bridge module projects Barbet's 1536-dimensional hidden states into the acoustic stack's input space, but the target dimension is not specified. Please state the acoustic stack's expected input dimension so readers can assess the bridge's capacity.
  6. §VII-E: The xRT (inverse real-time factor) is reported as median 4.75 and maximum 5.29, but the hardware is not specified. Please state the GPU or accelerator used, as this affects reproducibility of the speed claim.
  7. §VIII: The paper acknowledges two limitations (tokenizer-frontend coupling, short test sentences not exercising 262K context). These are appropriate disclosures. Consider also noting the single-speaker evaluation as a limitation, since all target-voice fine-tuning uses one voice.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the constructive assessment. The two major comments both concern the strength of evidence for the 25.2% frontend-isolated CER reduction: (1) potential ASR evaluator bias from using Breeze-ASR-25, and (2) the internal, unreleased test set. We address each below and commit to concrete revisions for both.

read point-by-point responses

  1. Referee: §VII-A, §VII-C: The 25.2% frontend-isolated CER reduction rests on the synthesize-then-recognize loop using Breeze-ASR-25, a Taiwan-localized model. Same-ASR does not guarantee unbiased comparison if the ASR has differential accuracy across acoustic regimes. If BlueMagpie-TTS produces speech whose accent, prosody, or code-switching timing more closely matches Breeze-ASR-25's training distribution, the ASR may transcribe BlueMagpie-TTS outputs more accurately than baseline outputs even if the actual intelligibility difference is smaller. The human preference study only compares BlueMagpie-TTS against the zero-shot base, not against the fine-tuned baseline that isolates the frontend effect.

    Authors: The referee raises a legitimate methodological concern. We agree that using the same ASR for all systems controls for absolute ASR difficulty but does not, by itself, rule out differential accuracy across acoustic regimes. If Breeze-ASR-25 is systematically better at transcribing Taiwanese-accented speech than other accents, and if BlueMagpie-TTS produces more Taiwanese-accented output, the ASR could favor our system for reasons unrelated to intelligibility. We acknowledge this as a genuine limitation of the synthesize-then-recognize methodology. We will take two steps in the revision. First, we will add an explicit discussion of this confound in Section VII-A, stating that same-ASR controls for absolute difficulty but not for differential accuracy, and noting the direction of the potential bias. Second, we will extend the human preference study to include the fine-tuned baseline (6.43% CER) as a third comparison arm, so that the frontend-isolated gap is validated by human listeners and not only by the ASR loop. We note that the existing human preference study already shows a strong effect (65.6% preference, p < 10^{-30}) for the zero-shot comparison, and the qualitative error analysis in Section VII-E (the base tends to drop or blur English spans, which the ASR then mis-recognizes) provides a mechanism that is not purely accent-matching: the base system fails to render English spans at all, which no ASR could transcribe correctly regardless of accent bias. Still, the referee's point stands for the 25.2% claim specifically, and the human study extension is the right fix. revision: yes

  2. Referee: §VII-B: The 1,000-sentence test set is internal and not publicly available. The central CER claim would be strengthened by releasing the test set, reporting on an external benchmark, or providing summary statistics (sentence length distribution, code-switching rate, English span types). Without this, it is unclear whether the 25.2% gap would hold on other Taiwan-localized test sets.

    Authors: This is a fair point. The headline result is specific to this test set, and without summary statistics or public availability, readers cannot assess generalizability. We will address this in two ways. First, we will add summary statistics for the test set in Section VII-B: sentence length distribution (mean, median, range), code-switching rate (fraction of sentences with at least one English span), and the breakdown of English span types (full words, abbreviations/acronyms, numbers/units, proper nouns), which we already categorize but do not currently report as counts. Second, we will release the test set publicly upon acceptance, subject to the licensing constraints of the source materials. We cannot report results on an existing external benchmark because, to our knowledge, there is no publicly available Taiwanese-Mandarin code-switching TTS benchmark with reference transcriptions. If the referee is aware of one, we would welcome the pointer. We will state the release plan explicitly in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity found: derivation chain is self-contained with external benchmarks and controlled comparisons

full rationale

The paper's three main claims each rest on independent evidence. (1) PangolinTokenizer's efficiency is measured on PangolinBench, which is author-created but transparent: it has explicit pass criteria, covers ten well-defined subsets, and compares against seven external tokenizers (GPT-4o, LLaMA 3.1, Qwen, Gemma, TAIDE, VoxCPM2). The vocabulary efficiency formula (Eq. 1) is explicitly stated as 'an empirical formula... not a derived quantity,' so it is not presented as a derivation. (2) Barbet's quality is evaluated on TAIDE-14 [7], an external benchmark from Taiwan's TAIDE project, against three external models (MiniCPM5-1B, LLaMA-3.2-1B, LFM2.5-1.2B). (3) The CER comparison uses Breeze-ASR-25 [29], an external ASR model from MediaTek Research, and the 25.2% frontend-isolated claim is supported by a controlled experimental design: the fine-tuned baseline and BlueMagpie-TTS share the same acoustic stack and same target-voice fine-tuning, differing only in the frontend and bridge. The paper discloses the limitation that 'the design couples the tokenizer and the frontend, so their contributions to CER are not separated' (Section VIII). No load-bearing self-citations exist: VoxCPM2 [22] is from OpenBMB, Breeze-ASR-25 [29] from MediaTek Research, TAIDE-14 [7] from the TAIDE Project. The reader's concern about potential ASR evaluator bias is a correctness risk, not circularity—the ASR is external and the same for all systems. The derivation chain does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The free parameters are tuning knobs (guidance strength, diffusion steps) applied uniformly across systems, so they do not bias the comparison but are still fitted. The load-bearing axiom is the ASR fairness assumption, which is unstated. The invented entities are engineering artifacts with empirical validation through the controlled ablation, not theoretical postulates.

free parameters (3)
  • Guidance strength = 2.8
    Tuned on the Taiwan-localized test set (Table VII); raised from default 2.0 to help the model commit to English tokens at switch points. Applied to all systems, so it does not favor one, but it is a fitted parameter.
  • Diffusion steps = 9
    Tuned on the test set; one below default of 10. Applied uniformly, but still a fitted parameter.
  • Bridge architecture (SwiGLU block) = 1 residual block, zero-initialized output
    Design choice for the bridge module; not justified by ablation against alternatives.
axioms (4)
  • domain assumption Breeze-ASR-25 is a fair and unbiased judge for comparing intelligibility across different TTS frontends
    Section VII-A: the synthesize-then-recognize loop uses this ASR for all systems. The paper states the ASR 'makes its own errors' but uses the same ASR for fairness. The assumption that the ASR's errors are not systematically biased toward one frontend is unstated but load-bearing for the CER comparison.
  • domain assumption The 1,000-sentence internal test set is representative of Taiwanese Mandarin code-switching usage
    Section VII-B: the test set is internal, not public, and is 'built from locally relevant cases on purpose.' The paper acknowledges it 'tests the accent and the switch points rather than providing an average-case measure.' Results on this set may not generalize to broader usage.
  • domain assumption VoxCPM2 acoustic stack is sufficiently strong that its frozen weights do not bottleneck quality
    Section III and VI: the design reuses the pretrained acoustic stack. If the acoustic stack were weak for Taiwanese Mandarin, the frontend improvement could be masked. The paper assumes the acoustic stack 'already produces high-fidelity audio.'
  • standard math Byte-level BPE is the appropriate tokenization method for Taiwan-context text
    Section IV-B: the lossless round-trip guarantee of byte-level BPE is a standard property. The choice of BPE over unigram or other schemes is a design decision, not proven optimal.
invented entities (2)
  • PangolinBench independent evidence
    purpose: Tokenizer benchmark with ten Taiwan-focused subsets for evaluating tokenizer suitability for Taiwan-localized TTS
    The benchmark has explicit pass criteria and is used to compare eight external tokenizers. Its contents are described by category. However, the full benchmark data is not clearly linked or released, which limits independent verification.
  • Bridge module (RMSNorm + linear projection + SwiGLU) independent evidence
    purpose: Maps Barbet hidden states into the input space of the VoxCPM2 acoustic stack
    The bridge is a standard projection module with a described training procedure (hidden-space distillation then joint fine-tuning). Its effectiveness is measured by the downstream CER. No new physics or entities are postulated; this is an engineering component.

pith-pipeline@v1.1.0-glm · 17451 in / 3361 out tokens · 449640 ms · 2026-07-08T17:50:38.431038+00:00 · methodology

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

35 extracted references · 35 canonical work pages · 15 internal anchors

  1. [1]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” inProc. International Conference on Machine Learning (ICML), 2021, pp. 5530–5540

    work page 2021

  2. [2]

    Seed-TTS: A Family of High-Quality Versatile Speech Generation Models

    P. Anastassiouet al., “Seed-TTS: A family of high-quality versatile speech generation models,”arXiv preprint arXiv:2406.02430, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

    Y . Chenet al., “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    V oxCPM: Tokenizer-free TTS for context-aware speech generation and true-to-life voice cloning,

    Y . Zhouet al., “V oxCPM: Tokenizer-free TTS for context-aware speech generation and true-to-life voice cloning,”arXiv preprint arXiv:2509.24650, 2025

    work page arXiv 2025

  5. [5]

    Neural machine translation of rare words with subword units,

    R. Sennrich, B. Haddow, and A. Birch, “Neural machine translation of rare words with subword units,” inProc. 54th Annual Meeting of the Association for Computational Linguistics (ACL), Berlin, Germany, 2016, pp. 1715–1725

    work page 2016

  6. [6]

    How good is your tokenizer? on the monolingual performance of multilingual language models,

    P. Rust, J. Pfeiffer, I. Vuli ´c, S. Ruder, and I. Gurevych, “How good is your tokenizer? on the monolingual performance of multilingual language models,” inProc. 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP), 2021, pp. 3118–3135

    work page 2021

  7. [7]

    TAIDE-14-tasks,

    TAIDE Project, “TAIDE-14-tasks,” https://huggingface.co/datasets/taide/ TAIDE-14-tasks, 2023

    work page 2023

  8. [8]

    Subword regularization: Improving neural network translation models with multiple subword candidates,

    T. Kudo, “Subword regularization: Improving neural network translation models with multiple subword candidates,” inProc. 56th Annual Meeting of the Association for Computational Linguistics (ACL), 2018, pp. 66–75

    work page 2018

  9. [9]

    SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,

    T. Kudo and J. Richardson, “SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,” inProc. Conference on Empirical Methods in Natural Language Processing (EMNLP): System Demonstrations, 2018, pp. 66–71

    work page 2018

  10. [10]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language models are unsupervised multitask learners,” OpenAI, Tech. Rep., 2019

    work page 2019

  11. [11]

    The Llama 3 Herd of Models

    A. Grattafioriet al., “The Llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Qwen2.5 Technical Report

    A. Yang, B. Yang, B. Zhanget al., “Qwen2.5 technical report,”arXiv preprint arXiv:2412.15115, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Gemma 3 Technical Report

    Gemma Team, “Gemma 3 technical report,” arXiv preprint arXiv:2503.19786, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Language model tokenizers introduce unfairness between languages,

    A. Petrov, E. L. Malfa, P. H. S. Torr, and A. Bibi, “Language model tokenizers introduce unfairness between languages,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  15. [15]

    Do all languages cost the same? tokenization in the era of commercial language models,

    O. Ahia, S. Kumar, H. Gonen, J. Kasai, D. R. Mortensen, N. A. Smith, and Y . Tsvetkov, “Do all languages cost the same? tokenization in the era of commercial language models,” inProc. Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023, pp. 9904– 9923

    work page 2023

  16. [16]

    Tokenization and the noiseless channel,

    V . Zouhar, C. Meister, J. L. Gastaldi, L. Du, M. Sachan, and R. Cotterell, “Tokenization and the noiseless channel,” inProc. 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023, pp. 5184– 5207

    work page 2023

  17. [17]

    ByT5: Towards a token-free future with pre-trained byte-to-byte models,

    L. Xue, A. Barua, N. Constant, R. Al-Rfou, S. Narang, M. Kale, A. Roberts, and C. Raffel, “ByT5: Towards a token-free future with pre-trained byte-to-byte models,”Transactions of the Association for Computational Linguistics, vol. 10, pp. 291–306, 2022

    work page 2022

  18. [18]

    CANINE: Pre- training an efficient tokenization-free encoder for language representation,

    J. H. Clark, D. Garrette, I. Turc, and J. Wieting, “CANINE: Pre- training an efficient tokenization-free encoder for language representation,” Transactions of the Association for Computational Linguistics, vol. 10, pp. 73–91, 2022

    work page 2022

  19. [19]

    Natural TTS synthesis by conditioning WaveNet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weisset al., “Natural TTS synthesis by conditioning WaveNet on mel spectrogram predictions,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 4779–4783

    work page 2018

  20. [20]

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

    Y . Ren, C. Hu, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y . Liu, “FastSpeech 2: Fast and high-quality end-to-end text to speech,”arXiv preprint arXiv:2006.04558, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  21. [21]

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

    C. Wanget al., “Neural codec language models are zero-shot text-to- speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    VoxCPM2 Technical Report

    Y . Zhouet al., “V oxCPM2 technical report,” OpenBMB, Tech. Rep., 2026, arXiv preprint arXiv:2606.06928

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model

    E. Casanovaet al., “XTTS: A massively multilingual zero-shot text-to- speech model,”arXiv preprint arXiv:2406.04904, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    End-to-end code- switching TTS with cross-lingual language model,

    X. Zhou, X. Tian, G. Lee, R. K. Das, and H. Li, “End-to-end code- switching TTS with cross-lingual language model,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7614–7618

    work page 2020

  25. [25]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems (NeurIPS), 2020

    work page 2020

  26. [26]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021

    work page 2021

  27. [27]

    Taiwan LLM: Bridging the Linguistic Divide with a Culturally Aligned Language Model

    Y .-T. Lin and Y .-N. Chen, “Taiwan LLM: Bridging the linguistic divide with a culturally aligned language model,”arXiv preprint arXiv:2311.17487, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    Measuring Taiwanese Mandarin Language Understanding

    P.-H. Chen, S. Cheng, W.-L. Chen, Y .-T. Lin, and Y .-N. Chen, “Mea- suring Taiwanese Mandarin language understanding,”arXiv preprint arXiv:2403.20180, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

    A Self-Refining Framework for Enhancing ASR Using TTS-Synthesized Data

    MediaTek Research, “Breeze-ASR-25: A Taiwanese-Mandarin and code- switching ASR model,”arXiv preprint arXiv:2506.11130, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inProc. International Conference on Machine Learning (ICML), 2023, pp. 28 492–28 518

    work page 2023

  31. [31]

    TAIDE: Trustworthy AI dialogue engine,

    TAIDE Project, “TAIDE: Trustworthy AI dialogue engine,” https://taide. tw, 2024, National Science and Technology Council, Taiwan

    work page 2024

  32. [32]

    MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies

    S. Huet al., “MiniCPM: Unveiling the potential of small language models with scalable training strategies,”arXiv preprint arXiv:2404.06395, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  33. [33]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  34. [34]

    GQA: Training generalized multi-query transformer models from multi-head checkpoints,

    J. Ainslie, J. Lee-Thorp, M. de Jong, Y . Zemlyanskiy, F. Lebrón, and S. Sanghai, “GQA: Training generalized multi-query transformer models from multi-head checkpoints,” inProc. Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023

    work page 2023

  35. [35]

    RoFormer: Enhanced Transformer with Rotary Position Embedding

    J. Su, Y . Lu, S. Pan, B. Wen, and Y . Liu, “RoFormer: Enhanced trans- former with rotary position embedding,”arXiv preprint arXiv:2104.09864, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021