Pith. sign in

REVIEW 3 major objections 4 minor 36 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a large language model can replace the weighted finite-state transducer stage in phoneme-based speech recognition, converting phoneme sequences into text and cutting word error rate by 3.6 percent on Polish and 6.9…

desk verdict A credible LLM-based P2G cascade with real but modest gains; the Polish headline is weakened by unclear hyperparameter selection, but the German result and the method's novelty justify peer review. read the letter →

arxiv 2506.04711 v1 pith:XNMZAYNY submitted 2025-06-05 cs.SD cs.CLeess.AS

Te Ma , Min Bi , Saierdaer Yusuyin , Hao Huang , Zhijian Ou This is my paper
classification cs.SDcs.CLeess.AS
keywords speechrecognitionphoneme-to-graphemelargelanguagemodeltop-KmarginalizedcrosslingualASRphoneme-baseddataaugmentationWFSTdecoding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes replacing the weighted finite-state transducer (WFST) decoding stage of phoneme-based speech recognition with a large language model, converting the phoneme stream produced by a speech-to-phoneme model into text. Its core claim is that this LLM-based phoneme-to-grapheme (LLM-P2G) cascade outperforms WFST decoding in crosslingual ASR when trained with two mitigation strategies: data augmentation with noisy phonemes (DANP) and randomized top-K marginalized (TKM) training and decoding. On 130 hours of Polish and German speech, the method reduces word error rate by 3.6% and 6.9% relative to WFST decoding with a language model, while simplifying the pipeline by removing lexicon construction and WFST compilation. If correct, the result means multilingual ASR can combine a phoneme-based acoustic model with an LLM's linguistic knowledge through a simple text-to-text interface, which matters for low-resource languages.

What carries the argument

The carrying mechanism is the latent-variable marginalization $p(y|x)\approx\sum_{h} p(h|x)p(y|h)$, where $h$ runs over a top-$K$ set of hypothesized phoneme sequences and $p(h|x)$ is computed by the CTC forward-backward algorithm. Randomized TKM training draws a random subset of $n$ of the top-$K$ hypotheses in each minibatch and maximizes the marginal likelihood, and TKM decoding scores candidate texts with the same formula, approximating $p(y|h)=0$ for text not generated by beam search from a hypothesis. This makes the LLM P2G model less sensitive to noisy S2P outputs. The companion mechanism, DANP, augments P2G training data with beam-search and sampled phoneme sequences from S2P, reducing the mismatch between clean training inputs and noisy test inputs. The conditional-independence factorization $p(y|x,h)=p(y|h)$ is the enabling equation, and the two strategies are what the paper uses to counter the information loss this equation can cause.

What would settle it

Run LLM-P2G with oracle (hand-verified) phoneme sequences instead of S2P hypotheses: if WER does not drop to near zero on homophone-heavy utterances, the conditional-independence assumption is losing information that no P2G model can restore, independent of S2P errors.

Watch

Extended reading notes

Core claim

The paper establishes that a two-step cascade, speech-to-phoneme (S2P) followed by LLM phoneme-to-grapheme (P2G), can serve as the decoding path for phoneme-based ASR. The architecture writes $p(y|x)=\sum_h p(h|x)p(y|h)$, treating the phoneme sequence $h$ as a latent variable. The authors argue that the naive version loses information because the S2P model's 1-best phoneme sequence is noisy and does not match clean training data. They fix this with DANP, which trains the LLM on noisy phoneme hypotheses, and with randomized TKM, which marginalizes over a random subset of top-K phoneme sequences in both training and decoding. With these techniques, LLM-P2G with randomized TKM beats WFST-based decoding under the same acoustic model: relative WER reductions of 3.6% for Polish (3.68 vs 3.82 with LM) and 6.9% for German (13.03 vs 14.01 with LM) at 130 hours. The same trend appears at 20 hours for German, while Polish low-resource does not improve, which the authors attribute to the LLM's pretraining data proportions.

Load-bearing premise

The load-bearing premise is the conditional-independence assumption that once the phoneme sequence is known, the audio adds no extra information about the text; any acoustic cue lost in the phoneme sequence, such as homophone disambiguation, word boundaries, stress, or speaking style, cannot be recovered by the phoneme-to-text stage.

Editorial extensions

If this is right

  • Phoneme-based ASR no longer requires lexicon construction and WFST compilation; decoding reduces to S2P beam search followed by LLM text generation.
  • LLM-P2G inherits the LLM's language priors: German, with more LLM pretraining data than Polish, benefits more, and at 20 hours only German improves over the phoneme baseline.
  • TKM decoding improves WER even when applied on top of a DANP-trained model, so marginalization over phoneme hypotheses contributes beyond data augmentation.
  • Because P2G treats phonemes and text as discrete tokens, the same architecture can plug in different LLMs or adapt to new languages by updating the P2G model alone.
  • The approach avoids projecting continuous speech embeddings into the LLM, using phonemes as a lighter interface between acoustic and language models.

Reading between the lines

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

  • A directly testable extension is to feed oracle phoneme sequences into the P2G model: any remaining WER isolates how much the conditional-independence assumption costs, independent of S2P errors.
  • Since the P2G stage is pure text transduction, zero-shot or few-shot P2G with a prompted LLM is a plausible next step for languages without fine-tuning data.
  • Randomized TKM could be applied to the S2P side as well, or to joint training of both stages, potentially closing more of the gap caused by information loss.
  • The Polish 20-hour result suggests that LLM-P2G's advantage is tied to LLM pretraining coverage, so target-language-specific pretraining or a larger multilingual LLM is a concrete lever to test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes replacing WFST-based decoding in phoneme-based ASR with a two-stage LLM-based phoneme-to-grapheme (LLM-P2G) decoder: Whistle-S2P converts speech to phoneme hypotheses, and a fine-tuned mT5-base converts phoneme sequences to subword text. To handle noisy or multiple hypothesized phoneme sequences, the authors propose data augmentation with noisy phonemes (DANP) and randomized top-K marginalized (TKM) training and decoding, which marginalize over subsets of S2P hypotheses. Experiments on Common Voice Polish and German at 130 and 20 hours compare LLM-P2G against Whistle phoneme/subword fine-tuning with WFST decoding. At 130 hours, randomized TKM gives 3.68 vs 3.82 WER for Polish and 13.03 vs 14.01 for German (relative reductions of 3.6% and 6.9%), with matched-pairs significance tests. Ablations in Tables 2-4 show consistent gains from DANP and TKM. The paper concludes that LLM-P2G outperforms WFST-based systems and simplifies the decoding pipeline.

Significance. If the 130-hour result is robust, the paper makes a useful contribution: it demonstrates that a fine-tuned LLM can serve as a phoneme-to-grapheme decoder, avoiding the construction of pronunciation lexicons and WFST graphs while leveraging LLM linguistic knowledge. The paper is commendable for releasing code, models, and data, using public datasets (Common Voice) and an external pretrained model (mT5), and for reporting matched-pairs significance tests and extensive ablations. The main caveat is that the headline claim is stated without qualification despite low-resource results that contradict it, and the significance of the best configuration is computed after selection from many informally explored settings, so the central claim needs additional validation evidence.

major comments (3)
  1. [Abstract; Section 5.1; Table 1] The abstract and Section 5.1 claim that LLM-P2G outperforms WFST-based systems in crosslingual ASR for Polish and German, with relative WER reductions of 3.6% and 6.9%. This is contradicted by the 20-hour rows in Table 1: for Polish, the best LLM-P2G row (randomized TKM, w/ LM) gives 17.36 versus 13.84 for Whistle Subword FT, i.e., about 25% worse; for German, it gives 28.78 versus 28.04. The claim should be restricted to the 130-hour setting, or the paper should provide a more substantive account of why the method fails in low-resource Polish and German rather than the brief mT5 pre-training data fraction explanation.
  2. [Section 4.2; Section 5.1; Tables 2-4] The headline comparison (Table 1, row 5) is the best of a large set of configurations selected without a described validation protocol. Table 2 reports seven DANP settings, Table 3 three TKM settings, Table 4 two training and two decoding strategies, and Section 4.2 mentions combining five checkpoints; the paper does not state which choices were fixed before test-time WERs were inspected or which split was used for selection. The matched-pairs p-values (p=1e-4 for Polish, p=8e-23 for German) are computed for the winning cell after selection, so they do not provide a valid post-selection significance level. Since the Polish gain is only 0.14 WER absolute (3.82 to 3.68), chance selection among roughly 10-15 configurations is a plausible alternative explanation. Please report validation-set WERs for all configurations tried, specify a selection rule, and either correct for multiple comparisons or compute p-values on a test set that was not used to choose the configuration.
  3. [Section 3.1; Eq. (1)] The two-step architecture relies on the conditional independence assumption p(y|x,h)=p(y|h). The paper identifies 'information loss in cascading S2P and P2G' and proposes DANP and TKM to mitigate it, but it never measures how much information about y is actually absent from h (e.g., homophones, word boundaries, stress, or speaking style). A concrete test would be to feed oracle or manually corrected phoneme sequences to the P2G model and compare with S2P 1-best input, or to compare the marginalized model of Eq. (1) against a model that additionally conditions on acoustic features. This would clarify what portion of the observed WER gap comes from the cascade assumption and what DANP/TKM can recover.
minor comments (4)
  1. [Section 5.2] The word 'randommized' should be 'randomized' in the sentence 'We train LLM-P2G using DANP or randommized TKM'.
  2. [Table 4] The column headers 'Best Path Decode' and 'TKM Decode' each have a 'w LM' column but no 'w/o LM' column, which makes the table confusing; add explicit 'w/o LM' columns or a note explaining the omitted condition.
  3. [Abstract; Section 3.3] The term 'top-Kmarginalized' should be 'top-K marginalized' with a space for readability.
  4. [Section 1; Section 3.1] The phrase 'there seems to have information loss' is awkward; suggest 'there appears to be information loss' or 'information is lost'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the cascade equations (1)-(3) are model definitions, and the WER comparisons are empirical measurements against external baselines.

full rationale

The paper's formal skeleton is Eqs. (1)-(3). Eq. (1) is an explicit conditional-independence assumption, p(y|x,h)=p(y|h), stated as an assumption for the two-step SPG architecture rather than as a theorem derived from data. Eq. (2) is a top-K approximation of that model definition, and Eq. (3) is a randomized variant of the training objective; none of these equations is claimed to follow from prior results or from the empirical findings. The central claim, that LLM-P2G with DANP and randomized TKM outperforms WFST decoding, is supported by WER measurements on the public Common Voice corpus using a fixed Whistle-S2P acoustic model, a publicly available mT5-base, and standard WFST baselines. The same Whistle-S2P is used for both the WFST and LLM-P2G rows, so the comparison isolates the P2G decoding contribution. The self-citations to Whistle [7] and CAT [23] provide the acoustic backbone and toolkit, but they are not used to argue for the P2G result; the experimental outcome is externally reproducible and is not derived from those citations. The DANP and TKM methods are evaluated through ablations, and no parameter fitted to the test set is renamed as a prediction. Although the matched-pairs p-value is reported for the best-performing row after several configurations were explored, that is a post-selection inference concern about statistical validity, not circularity: the reported WERs are not equivalent to their inputs by construction. Accordingly, no circular step is present.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends mainly on the conditional-independence assumption and on the noise-matching assumptions underlying DANP and TKM. Hyperparameters K, n, R, learning rate, and beam sizes are chosen by hand, and the best settings are selected from ablations. There are no invented entities.

free parameters (6)
  • K (number of top S2P phoneme hypotheses) = 32 for TKM training; 96-beam for best DANP; top-8 for TKM decoding
    Chosen by hand; the top-32 TKM setting performs badly (Polish WER 16.55) while top-8 works, so the result is sensitive to this choice.
  • n (number of randomly drawn hypotheses in randomized TKM) = 8
    Chosen by hand; not swept, and the best setting 'rand. 8 of top-32' depends on both K and n.
  • R (number of random S2P samples in DANP) = 25000 for Polish, 500 for German
    Chosen by hand; after de-duplication the augmented data grows about 32 times.
  • DANP beam size = 1, 32, 64, 96; 96-beam plus sampling is best
    Chosen by hand; the best setting also uses multiple checkpoints, so the reported result depends on several stacked choices.
  • mT5 fine-tuning learning rate = 3e-4
    Fixed learning rate with early stopping; no schedule or sweep is described.
  • P2G beam search size S and LM rescoring beam = 4
    Used in both WFST and LLM-P2G decoding; hand-set and not swept.
assumptions (4)
  • domain assumption Conditional independence p(y|x,h)=p(y|h)
    Invoked in Eq. (1), Section 3.1: text is conditionally independent of speech given phonemes. False for homophones or other acoustic-only distinctions.
  • domain assumption CTC-based sampling from S2P softmax approximates the true noise distribution of S2P hypotheses
    Section 3.2, DANP sampling relies on softmax probabilities being meaningful; if they are miscalibrated, augmentation may not match test-time errors.
  • domain assumption Top-K approximation with zero-filling for ungenerated sequences is accurate enough
    Section 3.4, decoding sets p(y|h(k)) approximately 0 when y was not generated by the beam from h(k), following RAG-Sequence Fast Decoding.
  • domain assumption mT5-base has sufficient pretraining coverage for Polish and German, including IPA tokens
    Section 1 and Section 5.1: the authors note that IPA symbols fall in mT5's token set, and later attribute German/Polish differences to pretraining data percentages (3.05% vs 2.15%).

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-based phoneme-to-grapheme for phoneme-based speech recognition." pith.science (2026). https://pith.science/paper/XNMZAYNY

@misc{pith2026250604711,
  author       = {Pith},
  title        = {Pith review of: LLM-based phoneme-to-grapheme for phoneme-based speech recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNMZAYNY}},
  note         = {Machine review of arXiv:2506.04711}
}
abstract

In automatic speech recognition (ASR), phoneme-based multilingual pre-training and crosslingual fine-tuning is attractive for its high data efficiency and competitive results compared to subword-based models. However, Weighted Finite State Transducer (WFST) based decoding is limited by its complex pipeline and inability to leverage large language models (LLMs). Therefore, we propose LLM-based phoneme-to-grapheme (LLM-P2G) decoding for phoneme-based ASR, consisting of speech-to-phoneme (S2P) and phoneme-to-grapheme (P2G). A challenge is that there seems to have information loss in cascading S2P and P2G. To address this challenge, we propose two training strategies: data augmentation with noisy phonemes (DANP), and randomized top-$K$ marginalized (TKM) training and decoding. Our experimental results show that LLM-P2G outperforms WFST-based systems in crosslingual ASR for Polish and German, by relative WER reductions of 3.6% and 6.9% respectively.

Figures

Figures reproduced from arXiv: 2506.04711 by the authors.

Figure 1
Figure 1. Phoneme-based ASR with WFST-based decoding (a) and with LLM-P2G decoding (b). In recognizing speech x into text y, phonemes arise as intermediate states, denoted by h. LLM-P2G decoding can be either best path decoding or top-K marginalized (TKM) decoding. In this work, we propose LLM-based phoneme-to￾grapheme (referred to as LLM-P2G) for phoneme-based ASR, as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

  1. [1]

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

    Introduction Most languages worldwide are under-resourced, posing sig- nificant challenges in developing high-performance ASR sys- tems. Therefore, multilingual pre-training and crosslingual fine-tuning have been developed, enabling information sharing and knowledge transferring between languages [1, 2, 3, 4, 5, 6]. Among these advancements, phoneme-based...

    work page Pith review arXiv 2025

  2. [2]

    These prior works share a similar motivation with ours that phoneme-based supervision is advantageous for multilingual acoustic representation learning

    Related Work The two-step idea of recognizing speech to phonemes and then to graphemes has been studied for crosslingual ASR [13, 14]. These prior works share a similar motivation with ours that phoneme-based supervision is advantageous for multilingual acoustic representation learning. However, both studies do not explore using LLMs for P2G for phoneme-b...

    work page

  3. [3]

    Whistle Subword FT

    Method 3.1. ASR architecture In the following, we introduce different ASR architectures, in- cluding the traditional one and the new architecture with LLM- P2G proposed in this work. For both architectures, the acoustic model (or say S2P) can be obtained by fine-tuning a phoneme- based multilingual S2P backbone (Whistle) over speech data with phoneme labe...

    work page

  4. [4]

    Dataset Experiments are conducted on the CommonV oice (CV) dataset [22], version 11.0 (released September 2022)

    Experiment 4.1. Dataset Experiments are conducted on the CommonV oice (CV) dataset [22], version 11.0 (released September 2022). Two languages from different language families, Polish (pl) and German (de), are selected, with 130 hours of training data per language, as they both use Latin script like in the pre-trained S2P model’s languages and are well-re...

    work page 2022

  5. [5]

    Results The main results are shown in Table 1

    Result and Ablation 5.1. Results The main results are shown in Table 1. For full training data (130 hours), the main observations are as follows: 1) LLM- P2G without DANP or TKM shows poor results, because of information loss (row 1 and 2 vs 3). 2) With DANP, for Pol- ish, while LLM-P2G does not surpass subword fine-tuning, it reduces WER by 2.7% compared...

    work page

  6. [6]

    Ablation results are shown in Table 3 and 4

    Top-8 is used in TKM decoding. Ablation results are shown in Table 3 and 4

    work page

  7. [7]

    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 Transactions on Audio, Speech and Language Processing, pp. 1–14, 2025

    work page 2025

  8. [8]

    First, the performance improves as the amount of augmented data increases. Second, the combined use of both beam search and random sampling yields superior results, compared to using Table 3:Word error rates (WERs) for LLM-P2G with different settings of TKM training and decoding. TKM strategy Polish German w/o LM w LM w/o LM w LM top-32 16.55 16.12 21.69 ...

    work page

Show all 36 references
  1. [9]

    Conclusion In this paper, we propose LLM-P2G for phoneme-based ASR, which belongs to a two-step ASR architecture, con- sisting of speech-to-phoneme and LLM-based phoneme-to- grapheme. Moreover, by incorporating data augmentation with noisy phonemes (DANP) and randomized top-Km...

  2. [10]

    Univer- sal phone recognition with a multilingual allophone system,

    X. Li, S. Dalmia, J. Li, M. Lee, P. Littell, J. Yao, A. Anastasopou- los, D. R. Mortensen, G. Neubig, A. W. Blacket al., “Univer- sal phone recognition with a multilingual allophone system,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Proc...

    2020

  3. [11]

    Multilingual and crosslin- gual speech recognition using phonological-vector based phone embeddings,

    C. Zhu, K. An, H. Zheng, and Z. Ou, “Multilingual and crosslin- gual speech recognition using phonological-vector based phone embeddings,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 1034–1041

    2021

  4. [12]

    Massively multilingual asr on 70 lan- guages: Tokenization, architecture, and generalization capabili- ties,

    A. Tjandra, N. Singhal, D. Zhang, O. Kalinli, A. Mohamed, D. Le, and M. L. Seltzer, “Massively multilingual asr on 70 lan- guages: Tokenization, architecture, and generalization capabili- ties,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal P...

    2023

  5. [13]

    Simple and effective zero-shot cross-lingual phoneme recognition,

    Q. Xu, A. Baevski, and M. Auli, “Simple and effective zero-shot cross-lingual phoneme recognition,” inINTERSPEECH, 2022, pp. 2113–2117

    2022

  6. [14]

    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,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  7. [15]

    Allophant: Cross- lingual phoneme recognition with articulatory attributes,

    K. Glocker, A. Herygers, and M. Georges, “Allophant: Cross- lingual phoneme recognition with articulatory attributes,” inIN- TERSPEECH, 2023, pp. 2258–2262

    2023

  8. [16]

    Can gener- ative large language models perform asr error correction?

    R. Ma, M. Qian, P. Manakul, M. Gales, and K. Knill, “Can gener- ative large language models perform asr error correction?”arXiv preprint arXiv:2307.04172, 2023

    2023 arXiv

  9. [17]

    Low-resourced speech recognition for iu mien language via weakly-supervised phoneme-based multilingual pretraining,

    L. Dong, D. Qin, F. Bai, F. Song, Y . Liu, C. Xu, and Z. Ou, “Low-resourced speech recognition for iu mien language via weakly-supervised phoneme-based multilingual pretraining,” in 2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP), 2024, pp...

    2024

  10. [18]

    Speech recognition with weighted finite-state transducers,

    M. Mohri, F. Pereira, and M. Riley, “Speech recognition with weighted finite-state transducers,” inSpringer Handbook of Speech Processing. Springer, 2008, pp. 559–584

    2008

  11. [19]

    mt5: A massively multilingual pre-trained text-to-text transformer,

    L. Xue, N. Constant, A. Roberts, M. Kale, R. Al-Rfou, A. Sid- dhant, A. Barua, and C. Raffel, “mt5: A massively multilingual pre-trained text-to-text transformer,” inProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguisti...

    2021

  12. [20]

    Whistle Phoneme FT

    on ten CV languages. We establish two baselines for each language by fine-tuning (FT) the Whistle-S backbone using weak phoneme labels 2 and subword labels, respectively. They are denoted by “Whistle Phoneme FT” and “Whistle Subword FT” in Table 1, respec- tively. The phoneme-...

  13. [21]

    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 preprint arXiv:2303.08774, 2023

    2023 arXiv

  14. [22]

    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 preprint arXiv:2302.13971, 2023

    2023 arXiv

  15. [23]

    TranUSR: Phoneme-to-word transcoder based unified speech representa- tion learning for cross-lingual speech recognition,

    H. Xue, Q. Shao, P. Chen, P. Guo, L. Xie, and J. Liu, “TranUSR: Phoneme-to-word transcoder based unified speech representa- tion learning for cross-lingual speech recognition,” inINTER- SPEECH, 2023, pp. 216–220

    2023

  16. [24]

    Optimizing two-pass cross- lingual transfer learning: Phoneme recognition and phoneme to grapheme translation,

    W. Lee, G. G. Lee, and Y . Kim, “Optimizing two-pass cross- lingual transfer learning: Phoneme recognition and phoneme to grapheme translation,” inIEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023

    2023

  17. [25]

    Emergent abilities of large language models,

    J. Wei, Y . Tay, R. Bommasani, C. Raffel, B. Zoph, S. Borgeaud, D. Yogatama, M. Bosma, D. Zhou, D. Metzler, E. H. Chi, T. Hashimoto, O. Vinyals, P. Liang, J. Dean, and W. Fedus, “Emergent abilities of large language models,”Transactions on Machine Learning Research, 2022

    2022

  18. [26]

    On decoder-only architecture for speech- to-text and large language model integration,

    J. Wu, Y . Gaur, Z. Chen, L. Zhou, Y . Zhu, T. Wang, J. Li, S. Liu, B. Ren, L. Liuet al., “On decoder-only architecture for speech- to-text and large language model integration,” in2023 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  19. [27]

    Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    2023 arXiv

  20. [28]

    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 preprint arXiv:2402.08846, 2024

    2024 arXiv

  21. [29]

    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,” inProceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

    2006

  22. [30]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K¨uttler, M. Lewis, W.-t. Yih, T. Rockt¨aschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Ad- vances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020

    2020

  23. [31]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), 2020, pp. 4211–4215

    2020

  24. [32]

    CAT: A CTC-CRF based ASR toolkit bridging the hybrid and the end-to-end approaches towards data efficiency and low latency,

    K. An, H. Xiang, and Z. Ou, “CAT: A CTC-CRF based ASR toolkit bridging the hybrid and the end-to-end approaches towards data efficiency and low latency,” inINTERSPEECH, 2020, pp. 566–570

    2020

  25. [33]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in INTERSPEECH, 2020, pp. 5036–5040

    2020

  26. [34]

    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 Language and Speech Processing, 2020. [Online]. Available: https://github.com/uiuc-sst/g2ps

    2020

  27. [35]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inProceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook, NY , USA: Curran Associates Inc., 2017, p. ...

    2017

  28. [36]

    Some statistical issues in the comparison of speech recognition algorithms,

    L. Gillick and S. J. Cox, “Some statistical issues in the comparison of speech recognition algorithms,” inInternational Conference on Acoustics, Speech, and Signal Processing (ICASSP), 1989

    1989

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.