arxiv: 2604.09332 · v1 · submitted 2026-04-10 · 📡 eess.AS

Recognition: unknown

Phonemes vs. Projectors: An Investigation of Speech-Language Interfaces for LLM-based ASR

Ziwei Li , Lukuang Dong , Saierdaer Yusuyin , Xianyu Zhao , Zhijian Ou

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:28 UTC · model grok-4.3

classification 📡 eess.AS

keywords ASRLLMphoneme interfaceprojectorspeech-language interfacelow-resource ASRBPE-phonemeTatar

0 comments

The pith

Phoneme-based interfaces outperform learned projectors for LLM-based ASR in low-resource settings.

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

The paper compares two approaches to linking pretrained speech encoders with large language models for automatic speech recognition. A learned projector maps continuous encoder features into the LLM embedding space, while a phoneme-based interface supplies discrete phoneme sequences instead. On high-resource English data the phoneme method is competitive and a BPE-phoneme variant that groups frequent local patterns improves results further. On low-resource Tatar the phoneme interface delivers substantially higher accuracy, and adding phoneme supervision produces a hybrid interface that beats the vanilla projector. This matters because interface design directly affects how much labeled speech data is needed to achieve usable performance, especially for languages with scarce resources.

Core claim

Using identical speech encoder and LLM backbones, the phoneme-based interface is competitive with the vanilla projector on LibriSpeech and substantially outperforms it on Tatar. The proposed BPE-phoneme interface, which groups frequent local phoneme patterns while preserving explicit word-boundary cues, yields further gains on English. Phoneme supervision produces a phoneme-informed hybrid interface that is stronger than the vanilla projector.

What carries the argument

The speech-language interface, implemented either as a learned projector mapping encoder features to LLM embeddings or as direct exposure of discrete phoneme sequences to the LLM.

Load-bearing premise

Phoneme labels or supervision must be available for training the interfaces, and identical encoder and LLM backbones must ensure a fair comparison.

What would settle it

A replication experiment on Tatar data showing no accuracy advantage for the phoneme-based interface over the vanilla projector would falsify the claim of substantial outperformance in low-resource conditions.

Figures

Figures reproduced from arXiv: 2604.09332 by Lukuang Dong, Saierdaer Yusuyin, Xianyu Zhao, Zhijian Ou, Ziwei Li.

**Figure 1.** Figure 1: Overview of the two types of speech–language interfaces studied in this work. (a) Projector-based: a trainable projector transforms downsampled speech-encoder ouput vectors into the LLM embedding space, followed by LoRA adaptation of the LLM. (b) Phoneme-based: we first train a CTC speech-to-phoneme (S2P) model, then feed sampled phoneme sequences into the LLM and adapt it with LoRA for phoneme-to-grapheme… view at source ↗

read the original abstract

Integrating pretrained speech encoders with large language models (LLMs) is promising for ASR, but performance and data efficiency depend on the speech-language interface. A common choice is a learned projector that maps encoder features into the LLM embedding space, whereas an alternative is to expose discrete phoneme sequences to the LLM. Using the same encoder and LLM backbones, we compare phoneme-based and vanilla projector-based interfaces in high-resource English and low-resource Tatar. We also propose a BPE-phoneme interface that groups frequent local phoneme patterns while preserving explicit word-boundary cues for phoneme-to-grapheme generation. On LibriSpeech, the phoneme-based interface is competitive with the vanilla projector, and the BPE-phoneme interface yields further gains. On Tatar, the phoneme-based interface substantially outperforms the vanilla projector. We further find that phoneme supervision yields a phoneme-informed hybrid interface that is stronger than the vanilla projector.

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.

Desk Editor's Note private letter to a colleague

Phoneme interfaces, especially the BPE variant, match or beat projectors in this controlled comparison, with the clearest gains on low-resource Tatar.

read the letter

The main takeaway is that phoneme-based interfaces are competitive with vanilla projectors on English and pull ahead on Tatar, and the new BPE-phoneme version adds measurable gains in both settings. They keep the speech encoder and LLM backbone identical across runs, so differences trace to the interface itself rather than model scale or pretraining tricks. On LibriSpeech the plain phoneme route holds even with the projector, the BPE grouping of frequent local patterns plus explicit boundaries improves further, and the hybrid that adds phoneme supervision is strongest. On Tatar the phoneme interface shows a substantial edge over the projector, and the hybrid keeps that advantage. This is useful because it tests the interface choice under both high- and low-resource conditions instead of staying English-only. The BPE proposal is a practical middle ground that keeps some linguistic structure without forcing the LLM to handle raw phoneme streams. The controlled backbone is the cleanest part of the work and makes the results easier to interpret. The soft spot is the Tatar supervision story. The abstract and stress-test note both flag that phoneme labels are used for the hybrid and the pure phoneme runs, but the paper needs to spell out whether those labels come from ground-truth transcripts, forced alignment on the same audio, or an external model that might have seen additional high-resource data. If the latter, the gap could partly reflect extra information rather than the interface design alone. A short methods paragraph or ablation on label source would close this. The work is aimed at people building or tuning LLM-based ASR pipelines, especially anyone working on under-resourced languages. It is worth sending to peer review because the question is practical, the comparison is set up to isolate the variable, and the low-resource results add something beyond standard benchmarks. A referee can ask for the missing details on supervision without needing major new experiments.

Referee Report

2 major / 2 minor

Summary. The manuscript compares phoneme-based and projector-based interfaces for connecting pretrained speech encoders to LLMs in ASR tasks. Using identical encoder and LLM backbones, it evaluates a vanilla projector, a direct phoneme interface, a proposed BPE-phoneme interface (grouping frequent phoneme patterns while preserving word boundaries), and a phoneme-informed hybrid on LibriSpeech (high-resource English) and a Tatar dataset (low-resource). Key findings are that phoneme-based methods are competitive on English with BPE-phoneme adding gains, while on Tatar the phoneme interface substantially outperforms the projector and phoneme supervision produces a stronger hybrid.

Significance. If the experimental conditions are shown to be equitable, the work offers useful empirical guidance on interface design for LLM-based ASR, particularly the value of explicit discrete phoneme representations in low-resource settings. The controlled use of shared backbones and the introduction of the BPE-phoneme grouping are strengths that allow direct attribution of differences to the interface choice.

major comments (2)

[Tatar experimental setup] Tatar experimental setup (likely §4 or Experiments): the manuscript must explicitly state how phoneme sequences or supervision signals are obtained for the Tatar data (ground-truth annotations, forced alignment on the identical audio, or an external phoneme model possibly trained on additional high-resource data). Without this, the headline claim that the phoneme-based interface substantially outperforms the vanilla projector cannot be evaluated for fairness, as the skeptic concern about informational advantage remains unaddressed.
[Tatar results table] Tatar results table (likely Table 2 or equivalent): the reported performance gaps lack error bars, standard deviations across runs, or statistical significance tests. This is load-bearing for the central claim of substantial outperformance and the hybrid advantage; add these to allow readers to assess whether the differences are reliable.

minor comments (2)

[Abstract] Abstract: the BPE-phoneme interface is introduced without a brief parenthetical definition of BPE in this context; add one sentence for readers unfamiliar with the term.
[Notation] Notation and terminology: ensure 'vanilla projector' is used consistently and defined on first use in the main text to avoid ambiguity with the hybrid variant.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the manuscript to improve clarity and strengthen the presentation of results.

read point-by-point responses

Referee: [Tatar experimental setup] Tatar experimental setup (likely §4 or Experiments): the manuscript must explicitly state how phoneme sequences or supervision signals are obtained for the Tatar data (ground-truth annotations, forced alignment on the identical audio, or an external phoneme model possibly trained on additional high-resource data). Without this, the headline claim that the phoneme-based interface substantially outperforms the vanilla projector cannot be evaluated for fairness, as the skeptic concern about informational advantage remains unaddressed.

Authors: We agree that explicit details on phoneme acquisition are required to allow fair evaluation of the results. We will revise the experimental setup section to clearly state that phoneme sequences for Tatar were obtained via forced alignment on the identical audio using a phoneme model trained solely on the provided Tatar training data (no additional high-resource language data was used for alignment beyond the shared pretrained encoder). This clarification directly addresses the concern about potential informational advantages and will be added in the next version. revision: yes
Referee: [Tatar results table] Tatar results table (likely Table 2 or equivalent): the reported performance gaps lack error bars, standard deviations across runs, or statistical significance tests. This is load-bearing for the central claim of substantial outperformance and the hybrid advantage; add these to allow readers to assess whether the differences are reliable.

Authors: We acknowledge that variability measures and significance tests would help readers evaluate the reliability of the reported gaps. Our experiments used single runs per configuration owing to the substantial computational requirements of LLM fine-tuning. In revision we will add standard deviations from a small number of additional runs with different random seeds (where already available) and include a brief discussion of consistency. Full statistical significance testing will be noted as a planned extension if time permits; otherwise we will qualify the claims accordingly. This constitutes a partial revision. revision: partial

Circularity Check

0 steps flagged

No circularity: pure empirical comparison

full rationale

The paper is an empirical study that trains and evaluates different speech-language interfaces (phoneme-based, BPE-phoneme, vanilla projector, and hybrid) on LibriSpeech and Tatar using fixed encoder and LLM backbones. All reported results are direct experimental outcomes (WER, etc.) rather than any derivation, first-principles prediction, or fitted-parameter renaming. No equations, self-citations as load-bearing premises, or ansatzes appear in the abstract or described content. The comparison is self-contained against external benchmarks and does not reduce to its own inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The claims rest on empirical results from experiments on LibriSpeech and Tatar datasets, assuming standard ML training practices and availability of phoneme annotations.

free parameters (1)

BPE vocabulary size
The size of BPE units for phonemes is likely chosen or tuned but not specified in abstract.

axioms (2)

domain assumption Discrete phoneme sequences can be directly input to LLMs for ASR tasks
This is the basis for the phoneme-based interface.
domain assumption Phoneme supervision during training improves the interface
Used to create the hybrid interface.

pith-pipeline@v0.9.0 · 5478 in / 1323 out tokens · 46871 ms · 2026-05-10T16:28:59.265808+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 17 canonical work pages · 6 internal anchors

[1]

Phonemes vs. Projectors: An Investigation of Speech-Language Interfaces for LLM-based ASR

Introduction Auto-regressive large language models (LLMs) have excelled in natural language processing (NLP) [1, 2], showing impressive language modeling and text generation capability. Exploring the potential of LLMs in automatic speech recognition (ASR) has therefore received increasing interest. Importantly, ASR is a cross-modality task, which is funda...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

Related work 2.1. End-to-End and Pretrained ASR Modern end-to-end ASR is commonly formulated with CTC [16], RNN-T [17], or attention-based encoder–decoder models [18], which unify acoustic modeling, alignment, and decoding in a single neural network. Large-scale speech pre- training further improves data efficiency: self-supervised ap- proaches such as wa...
[3]

Projector-based Interface As shown in Fig

Method 3.1. Projector-based Interface As shown in Fig. 1(a), the projector-interface approach con- sists of a speech encoder, a trainable projector, and an auto- regressive LLM decoder. For each utterance, we denote the speech signal asxand the target transcript asy. The speech encoder extracts frame-level acoustic representations: Hx = Enc(x),(1) whereH ...
[4]

Experimental Setup 4.1. Datasets We evaluate interface designs in both high-resource and low- resource settings.English (high-resource):We use Lib- riSpeech (960h) [33], validate ondev-other, and report WER ontest-cleanandtest-other.Tatar (low-resource):We use the Tatar (tt) subset of Common V oice [34] v11.0 (CV-tt, 20h). We validate ondevand report WER ...

2048
[5]

High-Resource Evaluation on LibriSpeech Table 1 reports LibriSpeech results under controlled back- bones (fixed LLM family and comparable LoRA adaptation)

Results 5.1. High-Resource Evaluation on LibriSpeech Table 1 reports LibriSpeech results under controlled back- bones (fixed LLM family and comparable LoRA adaptation). Vanilla projector baselines (E1–E2) are weak, suggesting that learning a robust continuous speech–text alignment from paired data alone can be challenging.Phoneme-informed hy- brid interfa...
[6]

Conclusion We compared projector-based and phoneme-based speech– language interfaces for LLM-ASR under controlled backbones. On LibriSpeech,vanillaprojectors with frozen encoders are weak; phoneme fine-tuning yields a strongphoneme-informed projector, while the best results are obtained by the BPE- phoneme interface. The BPE-phoneme gain in our imple- men...
[7]

They are not used to generate any core content, research ideas, experimental designs, results, or major textual parts of the paper

Generative AI Use Disclosure Generative AI tools are used in this work only for language edit- ing, polishing, and formatting of the manuscript. They are not used to generate any core content, research ideas, experimental designs, results, or major textual parts of the paper. All scien- tific contributions, including model design, experiments, analy- sis,...
[8]

Language models are few-shot learners,

T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” inProc. Advances in Neural Information Processing Systems (NeurIPS), 2020

2020
[9]

GPT-4 Technical Report

J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “GPT-4 technical report,”arXiv:2303.08774, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[10]

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

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inProc. Advances in Neural Information Processing Sys- tems (NeurIPS), 2020

2020
[11]

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

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

2021
[12]

WavLM: Large-scale self- supervised pre-training for full stack speech processing,

S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “WavLM: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

2022
[13]

Robust speech recognition via large-scale weak su- pervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inProc. International Conference on Machine Learn- ing (ICML), 2023

2023
[14]

Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision,

S. Yusuyin, T. Ma, H. Huang, W. Zhao, and Z. Ou, “Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision,”IEEE/ACM Trans. on Audio, Speech and Language Processing, vol. 33, pp. 1440–1453, 2025

2025
[15]

Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

W. Chan, N. Jaitly, Q. V . Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” inProc. International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2016

2016
[16]

Can gen- erative large language models perform ASR error correction?

R. Ma, M. Qian, P. Manakul, M. Gales, and K. Knill, “Can gen- erative large language models perform ASR error correction?” arXiv:2307.04172, 2023

work page arXiv 2023
[17]

LLM-based phoneme-to-grapheme for phoneme-based speech recognition,

T. Ma, M. Bi, S. Yusuyin, H. Huang, and Z. Ou, “LLM-based phoneme-to-grapheme for phoneme-based speech recognition,” in Proc. Interspeech, 2025

2025
[18]

Connecting speech encoder and large language model for ASR,

W. Yu, C. Tang, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “Connecting speech encoder and large language model for ASR,”arXiv:2309.13963, 2023

work page arXiv 2023
[19]

SALM: Speech-augmented language model with in-context learning for speech recognition and translation,

Z. Chen, H. Huang, A. Andrusenko, O. Hrinchuk, K. C. Puvvada, J. Li, S. Ghosh, J. Balam, and B. Ginsburg, “SALM: Speech- augmented language model with in-context learning for speech recognition and translation,”arXiv:2310.09424, 2023

work page arXiv 2023
[20]

An embarrassingly simple approach for llm with strong asr capacity,

Z. Ma, G. Yang, Y . Yang, Z. Gao, J. Wang, Z. Du, F. Yu, Q. Chen, S. Zheng, S. Zhanget al., “An embarrassingly simple approach for LLM with strong ASR capacity,”arXiv:2402.08846, 2024

work page arXiv 2024
[21]

A comprehensive solution to connect speech encoder and large language model for asr,

V . T. Pham, Y . Lin, T. Han, W. Li, J. Zhang, L. Lu, and Y . Wang, “A comprehensive solution to connect speech encoder and large language model for asr,”arXiv:2406.17272, 2024

work page arXiv 2024
[22]

A compara- tive study of discrete speech tokens for semantic- related tasks with large language models, 2024

D. Wang, M. Cui, D. Yang, X. Chen, and H. Meng, “A compar- ative study of discrete speech tokens for semantic-related tasks with large language models,”arXiv:2411.08742, 2024

work page arXiv 2024
[23]

Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” inProc. Interna- tional Conference on Machine Learning (ICML), 2006

2006
[24]

Sequence transduction with recurrent neural net- works,

A. Graves, “Sequence transduction with recurrent neural net- works,” inProc. International Conference on Machine Learning (ICML) Workshop on Representation Learning, 2012

2012
[25]

End-to- end continuous speech recognition using attention-based recurrent NN: First results,

J. Chorowski, D. Bahdanau, K. Cho, and Y . Bengio, “End-to- end continuous speech recognition using attention-based recurrent NN: First results,” inProc. NIPS Workshop on Deep Learning, 2014

2014
[26]

LLaMA: Open and Efficient Foundation Language Models

H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “LLaMA: Open and efficient foundation language models,” arXiv:2302.13971, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[27]

Qwen Technical Report

J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huanget al., “Qwen technical report,” arXiv:2309.16609, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[28]

Qwen3-ASR Technical Report

X. Shi, X. Wang, Z. Guo, Y . Wang, P. Zhang, X. Zhang, Z. Guo, H. Hao, Y . Xi, B. Yanget al., “Qwen3-ASR technical report,” arXiv:2601.21337, 2026

work page internal anchor Pith review arXiv 2026
[29]

OLMoASR: Open models and data for training robust speech recognition models,

H. Ngo, M. Deitke, M. Bartelds, S. Pratt, J. Gardner, M. Jordan, and L. Schmidt, “OLMoASR: Open models and data for training robust speech recognition models,”arXiv:2508.20869, 2025

work page arXiv 2025
[30]

Omnilingual asr: Open- source multilingual speech recognition for 1600+ languages.arXiv preprint arXiv:2511.09690, 2025

A. Omnilingual, G. Keren, A. Kozhevnikov, Y . Meng, C. Ropers, M. Setzler, S. Wang, I. Adebara, M. Auli, C. Baliogluet al., “Om- nilingual ASR: Open-source multilingual speech recognition for 1600+ languages,”arXiv:2511.09690, 2025

work page arXiv 2025
[31]

Gemini: A Family of Highly Capable Multimodal Models

G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millicanet al., “Gemini: a family of highly capable multimodal models,”arXiv:2312.11805, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[32]

AudioPaLM: A large language model that can speak and listen,

P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyen, A. Bapna, Z. Borsoset al., “AudioPaLM: A large language model that can speak and listen,”arXiv:2306.12925, 2023

work page arXiv 2023
[33]

I. P. Association,Handbook of the International Phonetic Associ- ation: A guide to the use of the International Phonetic Alphabet. Cambridge University Press, 1999

1999
[34]

Neural machine transla- tion of rare words with subword units,

R. Sennrich, B. Haddow, and A. Birch, “Neural machine transla- tion of rare words with subword units,” inProc. Annual Meeting of the Association for Computational Linguistics (ACL), 2016

2016
[35]

An investigation of phone-based subword units for end-to-end speech recognition,

W. Wang, G. Wang, A. Bhatnagar, Y . Zhou, C. Xiong, and R. Socher, “An investigation of phone-based subword units for end-to-end speech recognition,”arXiv:2004.04290, 2020

work page arXiv 2004
[36]

A systematic comparison of grapheme-based vs. phoneme-based label units for encoder-decoder-attention models,

M. Zeineldeen, A. Zeyer, R. Schlueter, and H. Ney, “A systematic comparison of grapheme-based vs. phoneme-based label units for encoder-decoder-attention models,”arXiv:2005.09336, 2021

work page arXiv 2005
[37]

Investigation into phone-based subword units for multilingual end-to-end speech recognition,

S. Yusuyin, H. Huang, J. Liu, and C. Liu, “Investigation into phone-based subword units for multilingual end-to-end speech recognition,” inProc. International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

2023
[38]

Phoneme-based speech recognition driven by large language models and sampling marginalization,

T. Ma, N. Li, H. Huang, and Z. Ou, “Phoneme-based speech recognition driven by large language models and sampling marginalization,” inProc. National Conference on Man-Machine Speech Communication (NCMMSC), 2025

2025
[39]

SentencePiece: A simple and lan- guage independent subword tokenizer and detokenizer for neural text processing,

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

2018
[40]

Lib- riSpeech: an ASR corpus based on public domain audio books,

V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- riSpeech: an ASR corpus based on public domain audio books,” inProc. International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2015

2015
[41]

Com- mon V oice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen- retty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Com- mon V oice: A massively-multilingual speech corpus,” inProc. International Conference on Language Resources and Evaluation (LREC), 2020

2020
[42]

LoRA: Low-rank adaptation of large language models,

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “LoRA: Low-rank adaptation of large language models,” inProc. International Conference on Learning Representations (ICLR), 2022

2022
[43]

Phonetisaurus: Ex- ploring grapheme-to-phoneme conversion with joint n-gram mod- els in the WFST framework,

J. R. Novak, N. Minematsu, and K. Hirose, “Phonetisaurus: Ex- ploring grapheme-to-phoneme conversion with joint n-gram mod- els in the WFST framework,”Natural Language Engineering, vol. 22, no. 6, pp. 907–938, 2016

2016
[44]

Grapheme-to-phoneme transduction for cross- language asr,

M. Hasegawa-Johnson, L. Rolston, C. Goudeseune, G.-A. Levow, and K. Kirchhoff, “Grapheme-to-phoneme transduction for cross- language asr,” inInternational Conference on Statistical Lan- guage and Speech Processing, 2020

2020
[45]

Pronunciation- lexicon free training for phoneme-based crosslingual ASR via joint stochastic approximation,

S. Yusuyin, T. Ma, H. Huang, and Z. Ou, “Pronunciation- lexicon free training for phoneme-based crosslingual ASR via joint stochastic approximation,”IEEE/ACM Trans. Audio, Speech, Language Process., vol. 34, pp. 272–284, 2026

2026