Pith. sign in

REVIEW 4 major objections 5 minor 48 references

LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context

T0 review · 4 major / 5 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 an LLM-based post-editor can fix rare-word ASR errors if fine-tuned on synthetic error pairs built from a rare-word list and conditioned on a simplified phonetic transcription, and reports lower WER/CER and sharp recall…

desk verdict A practical list-based GER recipe with a novel LSP phoneme feature, but the headline generalization claim is undercut by training and scoring on the same rare-word list. read the letter →

arxiv 2505.17410 v1 pith:WFROJZIG submitted 2025-05-23 cs.SD cs.CLeess.AS

Natsuo Yamashita , Masaaki Yamamoto , Hiroaki Kokubo , Yohei Kawaguchi This is my paper · ORCID
classification cs.SDcs.CLeess.AS
keywords automaticspeechrecognitiongenerativeerrorcorrectionlargelanguagemodelsrarewordssyntheticdatagenerationphoneticcontextN-besthypotheseskeywordbiasing
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 tries to establish that an LLM-based post-editor can reliably correct ASR mistakes on rare, domain-specific words when it is fine-tuned on error pairs synthesized from those words and when it is given the ASR's N-best hypotheses plus a simplified phonetic transcription of the 1-best output. The authors argue the two pieces do different work: the synthetic data teaches the model the target vocabulary and its typical confusions, while the phonetic context — an "LLM-based Simplified Phoneme" rendering — stops the model from over-correcting toward semantically plausible but phonetically wrong text. On five English and Japanese datasets the full method reports the best WER/CER in the paper, with rare-word recall rising from 27.6% to 85.0% on the Japanese medical-text set with ChatGPT. The reason to care is that rare words — names, places, technical terms — are the practical failure points of speech recognition, and the recipe needs nothing more than a word list and off-the-shelf generative components.

What carries the argument

The machinery is a two-stage pipeline. Stage one is synthetic error-pair generation: for each word in the rare-word list, the authors prompt an LLM for several example sentences, synthesize speech from them with multiple speakers using off-the-shelf TTS, and run a Whisper ASR model to get hypotheses; only pairs containing errors are kept for fine-tuning, so the model learns realistic recognition-error patterns for the target vocabulary. Stage two is LSP, the "LLM-based Simplified Phoneme": a plain-language phonetic reading of the ASR's 1-best output, generated by the LLM, appended to the N-best hypotheses as conditioning input. LSP is the object that carries the phonetic argument, because it is deliberately simpler than IPA, ARPAbet, or romanized Kana and therefore something the LLM can both produce and interpret.

What would settle it

Build a test set containing rare words deliberately excluded from the rare-word list used to synthesize fine-tuning data, and measure rare-word recall on those held-out words; if recall is no better than the un-corrected ASR baseline, the reported gains come from memorizing the list rather than from a general correction ability. A cheaper check is to fine-tune on one domain's rare-word list and evaluate on another domain's rare words.

Watch

Extended reading notes

Core claim

The central claim is that generative error correction for ASR is not fundamentally limited by scarce rare-word training data, because the needed error pairs can be synthesized: an LLM writes varied sentences containing each rare word, a text-to-speech model speaks them with multiple voices, and an ASR model transcribes them to produce aligned correct/error pairs for fine-tuning. The second claim is that phonetic context should be supplied as a simplified phoneme string written by the LLM itself, rather than IPA or standard TTS phonemes, because the simplified form is easier for the LLM to align with the text and prevents over-correction. In the paper's experiments these two components together outperform prompt-only and N-best-only baselines, and the simplified phoneme input yields further WER/CER gains across all evaluated datasets, in contrast to IPA, which sharply degrades Japanese CER.

Load-bearing premise

The load-bearing premise is that training on synthetic errors for a rare word teaches the model to correct that same word when it appears in real speech, and the paper's evaluation builds the rare-word list from each test set and then uses exactly that list to synthesize the fine-tuning data, so the measured recall gains could reflect memorization of the supplied list rather than general correction of unseen rare words.

Editorial extensions

If this is right

  • An ASR pipeline could be adapted to a new domain's vocabulary with only a rare-word list, removing the need to collect human-labeled error corpora for every new set of terms.
  • Because the phonetic context is derived from text rather than from internal acoustic features, the method stays applicable even when the ASR model is a closed API that only exposes N-best hypotheses.
  • The synthetic training distribution is generated by the same kinds of LLM and TTS components the method already uses, so the recipe should scale as those components improve.
  • On the paper's numbers, the gains are largest where homophones and complex terms are common, suggesting the approach targets exactly the failure modes standard GER baselines leave unresolved.

Reading between the lines

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

  • Editorial inference: the decisive untested question is whether the model corrects rare words it has never seen during fine-tuning; the current evaluation scores recall on the same word list used to build the training data, so a held-out rare-word split is needed to separate list memorization from true correction.
  • Editorial inference: the LSP design implies a cheap practitioner experiment — compare a fine-tuned LLM given N-best text alone against the same model given those hypotheses plus a naive phonemic spelling — and the paper's tables suggest the phoneme channel earns a small but consistent WER/CER margin.
  • Editorial inference: because the training speech is synthesized from text, the method inherits a TTS-to-real-speech gap; a natural next test is noisy or strongly accented audio, where phonetic context should matter most and where synthetic error pairs may be least representative.
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

4 major / 5 minor

Summary. The paper proposes a generative error correction (GER) pipeline that first synthesizes fine-tuning data from a rare-word list (LLM-generated transcripts, TTS speech, then Whisper error pairs), then fine-tunes an LLM on the ASR N-best hypotheses together with a simplified phonetic context called LSP. The method is evaluated on five English and Japanese datasets (LibriSpeech, EDGAR, CSJ eval1/eval2, MedTxt) with ChatGPT-4o-mini and Llama-3.3-70B variants, reporting WER/CER, rare-word recall, and precision. The authors claim that the proposed method improves rare-word correction and reduces WER/CER across all datasets.

Significance. If the reported results reflect genuine generalization, the proposed pipeline is practically valuable: it gives a concrete recipe for adapting LLM-based GER to domains with rare words using only a rare-word list, and the LSP representation is a simple way to inject phonetic information without external G2P tools. The manuscript also has reproducible components: the rare-word lists are publicly released, and the experimental configuration (T=4 transcripts, S=7 speakers, N=5 hypotheses, LoRA rank 16) is clearly stated. However, the current evaluation does not establish that the model corrects rare words it has not seen during fine-tuning, because the same rare-word lists are used both to synthesize training data and to score test recall. The significance therefore depends on an evaluation redesign that separates training and test rare words.

major comments (4)
  1. [§4.2–4.3, Tables 3–4] The central evaluation is compromised by a direct train/test overlap. Section 4.2 states that rare-word lists were constructed for LibriSpeech, EDGAR, and CSJ, and Section 4.3 states that synthetic transcripts and utterances were generated from those lists and used to fine-tune the GER model. Tables 3 and 4 then report rare-word recall over the same lists. Under this design, a large recall gain (e.g., MedTxt from 27.6% to 85.0% with ChatGPT) is consistent with the model memorizing the fine-tuning list rather than with an ability to correct unseen rare words. The paper needs a held-out rare-word evaluation, for example a second rare-word list that is not used to synthesize training data, or an explicitly framed deployment setting where the test list is available at fine-tuning time. The future-work sentence in Section 6, which says generalizability remains to be assessed, effectively concedes that this question is open; it should be resolved before the central claim is made.
  2. [§5.1, Tables 3–4] No confidence intervals, significance tests, or utterance-level variance are reported, and several headline differences are small. For example, LibriSpeech WER improves from 2.7% (Whisper) to 2.5% (Synth+N-best+LSP) with ChatGPT, and CSJ eval1 CER improves from 14.2% to 13.7%; with roughly 2600 test utterances on LibriSpeech and 1400 on CSJ, these differences may be within noise. Please report bootstrap confidence intervals or paired significance tests over test utterances, and state the number of rare-word tokens scored. This is needed to support the abstract's claim that the method 'reduces the WER and CER across both English and Japanese datasets.'
  3. [§3.2 and Table 2] The LSP contribution is underspecified. Section 3.2 says LSP is generated by prompting an LLM to 'convert text to simplified pronunciation,' but the exact prompt, the output format, and any verification of LSP correctness are not described. Without this, it is difficult to reproduce the method or to know whether the reported LSP gains are due to the phonetic representation itself or to the particular prompt used. Please provide the full LSP prompt templates, give examples for Japanese, and report a small accuracy or consistency check of LSP predictions on the test words.
  4. [§4.1, §4.3, Table 2] The nature of the EDGAR and MedTxt test sets needs clarification. Table 2 shows that EDGAR and MedTxt evaluation audio is synthesized (VITS trained on VCTK; FastSpeech2/HiFi-GAN trained on JSUT), while Section 4.3 says synthetic Japanese training data was produced with Microsoft Azure TTS. The paper should state explicitly which TTS voices and text prompts are used for training versus test utterances for these two datasets, and explain how any overlap between training-time synthetic audio and test-time synthetic audio is avoided. Otherwise the WER/CER gains on EDGAR and MedTxt may partly reflect matched TTS conditions rather than generalizable correction.
minor comments (5)
  1. [§3.1] The first sentence contains the typo 'Wbiasning words'; it should be 'biasing words.'
  2. [Table 2] For LibriSpeech and CSJ, the TTS model and TTS train-data columns contain dashes, but the method requires synthetic speech for fine-tuning. Please specify which TTS system was used to generate the synthetic training utterances for these two datasets.
  3. [Figure 2 and §5.2] The dataset name is written inconsistently as 'Medtxt' in Figure 2's caption and as 'MedTxt' elsewhere; please make the capitalization uniform.
  4. [References] Reference [17] has the typo 'Interspcch' and should be 'Interspeech'; also, the reference list would benefit from including page numbers or article numbers for these conference papers where available.
  5. [§4.2] The rare-word lists are constructed by asking an LLM to extract 'highly complex words,' but the paper does not report any agreement measure or manual verification of these lists. Given that the lists are a core input to the method, a short description of their size and a few examples for each dataset would improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Rare-word recall is evaluated on the same rare-word list used to generate the synthetic fine-tuning data, so the headline recall gains are not an independent prediction.

  1. fitted input called prediction [Sections 3.1, 4.2, 4.3, 4.6; Tables 3 and 4]
    "Given a set of biasing words {w_n}, we generate T synthesized transcripts ... using an ASR model, we obtain W×T×S hypotheses paired with the corresponding transcripts to fine-tune a GER model. ... we constructed rare word lists for each dataset with the assistance of an LLM prompt ... we reported the recall and the precision of the rare words, following the methodology described in [40]."

    The 'biasing words' used to generate the synthetic fine-tuning data in Section 4.3 are the same rare-word lists that Section 4.2 constructs from each test set for LibriSpeech, EDGAR, and CSJ, and the same style of list is used for the other datasets. The recall metric in Section 4.6 then scores exactly those words in Tables 3 and 4, so every rare word in the evaluation has already appeared in the model's synthetic training data, often in multiple transcripts and speakers. The headline gains (e.g., MedTxt recall from 27.6% to 85.0% with ChatGPT) can therefore be produced by memorizing the supplied list rather than by a generalizable correction capability. The WER and CER reductions are also partly confounded, because test utterances contain the same trained rare words.

full rationale

The paper's central empirical claim is that the proposed GER pipeline improves rare-word correction and reduces WER/CER. The evidence for the rare-word recall part is circular in the evaluation-design sense: the rare-word list is both the source of synthetic fine-tuning data and the scoring target. For four of the five test sets, the list is explicitly constructed from the test set itself, so the model is trained on the exact word types it is later asked to correct. This does not establish the ability to correct unseen rare words, and the paper's own future-work statement acknowledges that generalizability remains open. The whole-utterance WER/CER improvements are somewhat more independent, because they involve the entire utterance rather than isolated list words, and the comparison among phonetic representations (IPA, TTS-phoneme, LSP) is an internally coherent ablation that is not itself circular. No load-bearing self-citation chain or imported uniqueness argument appears in the paper. Taking these factors together, the evaluation loop affects the central rare-word claim and partially contaminates the WER/CER claim, but the method still has independently testable components; hence a score of 6 rather than 8 or 10.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim depends on several hand-chosen hyperparameters and on domain assumptions about synthetic error realism, LLM-generated phoneme quality, and the validity of using the same rare-word list for training and evaluation. These are not derived constants or externally benchmarked facts.

free parameters (8)
  • T (synthetic transcripts per rare word) = 4 per rare word
    Chosen by hand to balance diversity and compute; Figure 2 shows improvement from 1 to 4 transcripts, so it is an empirical knob, not a derived constant.
  • S (TTS speakers per transcript) = 7 speakers
    Chosen by hand; Figure 2 shows speaker diversity has a smaller effect than transcript count.
  • N (number of ASR N-best hypotheses) = 5
    Set for beam search from Whisper; affects input context and computational cost.
  • LoRA rank = 16
    Applied to all linear modules; chosen without reported tuning.
  • ChatGPT fine-tuning epochs, batch size, LR multiplier = 3 epochs, batch 10, LR multiplier 1.8
    Hand-selected for fine-tuning ChatGPT-4o-mini.
  • Llama fine-tuning epochs, batch size, learning rate = 10 epochs, batch 4, LR 5e-5
    Hand-selected for LoRA fine-tuning of Llama models on H100 GPUs.
  • Synthetic data validation split ratio = 4:1 train/validation
    Chosen to hold out 20% of synthetic pairs; no sensitivity analysis is reported.
  • Rare word coverage cap = <10% of total words
    Used to guide rare word list construction per prior work; affects which words are labeled rare.
assumptions (5)
  • domain assumption Synthetic error-pair data generated from LLM transcripts, TTS audio, and Whisper hypotheses are representative of real ASR error patterns.
    Section 3.1 builds the entire fine-tuning set this way; if synthetic errors do not match real errors, the transfer fails.
  • domain assumption Rare word lists extracted by an LLM from each test dataset are accurate and are the correct target words.
    Section 4.2 uses LLM prompts to build lists; errors in the lists change both training and evaluation.
  • domain assumption An LLM can produce simplified phonetic readings (LSP) accurate enough for correction, despite known LLM difficulties with G2P (cited [25]).
    Section 3.2 introduces LSP with no independent accuracy evaluation; the method assumes the LLM-generated phonemes are reliable.
  • domain assumption Whisper's N-best hypotheses and 1-best phonemes contain enough information for the GER model to correct errors without access to audio.
    Sections 4.3 and 4.4 set N=5 and use only 1-best phonemes to limit cost; if errors are absent from all N-best hypotheses, correction must come from the LLM's prior.
  • ad hoc to paper The same rare-word list may be used for synthetic training and test scoring without invalidating the comparison.
    Sections 4.2 and 4.3 use the same lists for data generation and evaluation; this is specific to this paper's protocol and is not independently justified.
invented entities (1)
  • LLM-based Simplified Phoneme (LSP)
    purpose: A simplified phonetic spelling of the 1-best ASR hypothesis, fed to the GER model alongside N-best text to reduce over-correction.
    LSP is introduced in Section 3.2 and evaluated only within this pipeline. No external benchmark validates its accuracy or shows it is a stable representation; the comparison to IPA and TTS-phonemes in Tables 3 and 4 is the only evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context." pith.science (2026). https://pith.science/paper/WFROJZIG

@misc{pith2026250517410,
  author       = {Pith},
  title        = {Pith review of: LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFROJZIG}},
  note         = {Machine review of arXiv:2505.17410}
}
read the original abstract

Generative error correction (GER) with large language models (LLMs) has emerged as an effective post-processing approach to improve automatic speech recognition (ASR) performance. However, it often struggles with rare or domain-specific words due to limited training data. Furthermore, existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction. To address these issues, we propose a novel LLM-based GER approach that targets rare words and incorporates phonetic information. First, we generate synthetic data to contain rare words for fine-tuning the GER model. Second, we integrate ASR's N-best hypotheses along with phonetic context to mitigate over-correction. Experimental results show that our method not only improves the correction of rare words but also reduces the WER and CER across both English and Japanese datasets.

Figures

Figures reproduced from arXiv: 2505.17410 by the authors.

Figure 1
Figure 1. Overview of the proposed methods of synthetic data generation from rare words and GER with phonetic context. 2.2. Phonetic notation In the field of linguistics, the International Phonetic Alpha￾bet (IPA) [21] has been widely used to accurately represent the pronunciation of all languages around the world. How￾ever, it is often perceived as complex, particularly by non-native speakers, due to the specialized linguist… view at source ↗
Figure 2
Figure 2. illustrates how the number of transcripts and speakers during synthetic data generation impacts the rare word recogni￾tion performance in the MedTxt dataset when using ChatGPT. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 40 canonical work pages

  1. [1]

    However, these systems often produce transcription errors, particularly due to back- ground noise, speaker accents, different speaker styles, and domain-specific terms

    Introduction Automatic speech recognition (ASR) technology has achieved remarkable progress over the past decades, driven by advance- ments in deep learning techniques [1]. However, these systems often produce transcription errors, particularly due to back- ground noise, speaker accents, different speaker styles, and domain-specific terms. To address this...

    work page

  2. [2]

    Related Work 2.1. Error correction for rare words In many practical scenarios, a rare word list can be obtained in advance from sources such as user names, meeting chat logs, websites, manuals, or even words registered by users [17]. Sev- eral studies [18, 19] have explored the use of rare word lists for error correction, but the error-pair data is limite...

    work page Pith review arXiv 2025

  3. [3]

    Provide 5 dif- ferent English sentences in various contexts that include the termw n, which is a medical term

    Methodology Figure 1 illustrates our proposed methods. First, we propose training a GER model using synthetic data generated by gera- tive models based on a given rare word list. Additionally, to prevent over-correction caused by ignoring phonetic informa- tion, we propose explicitly providing the GER model not only N-best hypotheses, but also 1-best phon...

    work page

  4. [4]

    Extract highly complex words for recognition, including tech- nical terms, names of people, and names of places

    Experimental Setup 4.1. Dataset To investigate the efficiency of our approach in a multilingual setting, we evaluated both English and Japanese datasets, as summarized in Table 2. Since public datasets like Common V oice [26] typically contain very few rare words and are of- ten included in the training sets of foundation models, we con- structed syntheti...

    work page 1993

  5. [5]

    WER / recall / precision

    Results and Analysis 5.1. Results Table 3 and 4 present the WER for the English datasets, the CER for the Japanese datasets, as well as the recall and pre- cision scores when using ChatGPT or the Llama models as the LLMs. First, we note that most of the results using Chat- GPT were better than those using the LLama models. The ini- tial transcription resu...

    work page

  6. [6]

    We intro- duced a method for generating diverse synthetic data contain- ing rare words, combined with leveraging LLM-based simpli- fied phonemes to avoid over-correction

    Conclusions In this paper, we proposed an approach to enhance the LLM- based GER for transcripts containing rare words. We intro- duced a method for generating diverse synthetic data contain- ing rare words, combined with leveraging LLM-based simpli- fied phonemes to avoid over-correction. The experimental re- sults demonstrated that our approach achieved...

    work page

  7. [7]

    Generative error correction for code-switching speech recognition using large language models

    C. Chen, Y . Hu, C.-H. H. Yang, H. Liu, S. M. Siniscalchi, and E. S. Chng, “Generative error correction for code-switching speech recognition using large language models,”arXiv preprint arXiv:2310.13013, 2023

    work page Pith review arXiv 2023

  8. [8]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl¨uter, and S. Watan- abe, “End-to-end speech recognition: A survey,” inIEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    work page 2024

Show all 48 references
  1. [9]

    Non-autoregressive error correction for CTC-based ASR with phone-conditioned masked LM,

    H. Futami, H. Inaguma, S. Ueno, M. Mimura, S. Sakai, and T. Kawahara, “Non-autoregressive error correction for CTC-based ASR with phone-conditioned masked LM,” inProc. Interspeech, 2022, pp. 3889–3893

    2022

  2. [10]

    Spelling error correction with soft-masked BERT,

    S. Zhang, H. Huang, J. Liu, and H. Li, “Spelling error correction with soft-masked BERT,” inProc. ACL, 2020, pp. 882–890

    2020

  3. [11]

    Distilling the knowledge of BERT for sequence- to-sequence asr,

    H. Futami, H. Inaguma, S. Ueno, M. Mimura, S. Sakai, and T. Kawahara, “Distilling the knowledge of BERT for sequence- to-sequence asr,” inProc. Interspeech, 2020, pp. 3635–3639

    2020

  4. [12]

    Investigating asr error correction with large language model and multilingual 1-best hypotheses,

    S. Li, C. Chen, C. Y . Kwok, C. Chu, E. S. Chng, and H. Kawai, “Investigating asr error correction with large language model and multilingual 1-best hypotheses,” inProc. Interspeech, 2024, pp. 1315–1319

    2024

  5. [13]

    Multi-stage large lan- guage model correction for speech recognition,

    J. Pu, T.-S. Nguyen, and S. St ¨uker, “Multi-stage large lan- guage model correction for speech recognition,”arXiv preprint arXiv:2310.11532, 2023

    2023 arXiv

  6. [14]

    It’s never too late: Fusing acoustic informa- tion into large language models for automatic speech recognition,

    C. CHEN, R. Li, Y . Hu, S. M. Siniscalchi, P.-Y . Chen, E. Chng, and C.-H. H. Yang, “It’s never too late: Fusing acoustic informa- tion into large language models for automatic speech recognition,” inProc. ICLR, 2024, pp. 1–17

    2024

  7. [15]

    Hyporadise: An open baseline for generative speech recognition with large language models,

    C. Chen, Y . Hu, C.-H. H. Yang, S. M. Siniscalchi, P.-Y . Chen, and E.-S. Chng, “Hyporadise: An open baseline for generative speech recognition with large language models,”Advances in Neural In- formation Processing Systems, vol. 36, 2024

    2024

  8. [16]

    N-best T5: Ro- bust asr error correction using multiple input hypotheses and con- strained decoding space,

    R. Ma, M. J. Gales, K. M. Knill, and M. Qian, “N-best T5: Ro- bust asr error correction using multiple input hypotheses and con- strained decoding space,” pp. 3267–3271, 2023

    2023

  9. [17]

    Bench- marking Japanese speech recognition on ASR-LLM setups with multi-pass augmented generative error correction,

    Y . Ko, S. Li, C.-H. H. Yang, and T. Kawahara, “Bench- marking Japanese speech recognition on ASR-LLM setups with multi-pass augmented generative error correction,”arXiv preprint arXiv:2408.16180, 2024

    2024 arXiv

  10. [18]

    Generative speech recognition error correction with large language models and task-activating prompting,

    C.-H. H. Yang, Y . Gu, Y .-C. Liu, S. Ghosh, I. Bulyko, and A. Stol- cke, “Generative speech recognition error correction with large language models and task-activating prompting,” inProc. ASRU, 2023, pp. 1–8

    2023

  11. [19]

    Dictionary- based phrase-level prompting of large language models for ma- chine translation,

    M. Ghazvininejad, H. Gonen, and L. Zettlemoyer, “Dictionary- based phrase-level prompting of large language models for ma- chine translation,”arXiv preprint arXiv:2302.07856, 2023

    2023 arXiv

  12. [20]

    Can large language models understand uncommon mean- ings of common words?

    J. Wu, F. Che, X. Zheng, S. Zhang, R. Jin, S. Nie, P. Shao, and J. Tao, “Can large language models understand uncommon mean- ings of common words?” 2024

    2024

  13. [21]

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

    1999

  14. [22]

    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

  15. [23]

    ChatGPT

    OpenAI, “ChatGPT.” [Online]. Available: https://openai.com

  16. [24]

    InterBiasing: Boost unseen word recognition through biasing intermediate predictions,

    Y . Nakagome and M. Hentschel, “InterBiasing: Boost unseen word recognition through biasing intermediate predictions,” in Proc. Interspcch, 2024, pp. 207–211

    2024

  17. [25]

    ED-CEC: Improving rare word recognition using asr postprocessing based on error detection and context-aware error correction,

    J. He, Z. Yang, and T. Toda, “ED-CEC: Improving rare word recognition using asr postprocessing based on error detection and context-aware error correction,” inProc. ASRU, 2023, pp. 1–6

    2023

  18. [26]

    Entity resolution for noisy ASR transcripts,

    A. Raghuvanshi, V . Ramakrishnan, V . Embar, L. Carroll, and K. Raghunathan, “Entity resolution for noisy ASR transcripts,” inProc. EMNLP-IJCNLP, 2019, pp. 61–66

    2019

  19. [27]

    Retrieval augmented correction of named entity speech recognition errors,

    E. Pusateri, A. Walia, A. Kashi, B. Bandyopadhyay, N. Hyder, S. Mahinder, R. Anantha, D. Liu, and S. Gondala, “Retrieval augmented correction of named entity speech recognition errors,” arXiv preprint arXiv:2409.06062, 2024

    2024 arXiv

  20. [28]

    EDGAR-CORPUS: Billions of tokens make the world go round,

    L. Loukas, M. Fergadiotis, I. Androutsopoulos, and P. Malakasi- otis, “EDGAR-CORPUS: Billions of tokens make the world go round,” inProc. ECONLP, 2021, pp. 13–18

    2021

  21. [29]

    Simplified Japanese pho- netic alphabet as a tool for Japanese course design,

    M. Kawai, S. Akai, and R. Shirakawa, “Simplified Japanese pho- netic alphabet as a tool for Japanese course design,” inProc. CA- JLE, 2019, pp. 146–155

    2019

  22. [30]

    Parallel Tacotron 2: A non-autoregressive neural TTS model with differentiable duration modeling,

    I. Elias, H. Zen, J. Shen, Y . Zhang, Y . Jia, R. Skerry-Ryan, and Y . Wu, “Parallel Tacotron 2: A non-autoregressive neural TTS model with differentiable duration modeling,” 2021

    2021

  23. [31]

    Data driven grapheme-to-phoneme representations for a lexicon-free text-to- speech,

    A. Garg, J. Kim, S. Khyalia, C. Kim, and D. Gowda, “Data driven grapheme-to-phoneme representations for a lexicon-free text-to- speech,” inProc. ICASSP, 2024, pp. 11 091–11 095

    2024

  24. [32]

    LLM- powered grapheme-to-phoneme conversion: Benchmark and case study,

    M. F. Qharabagh, Z. Dehghanian, and H. R. Rabiee, “LLM- powered grapheme-to-phoneme conversion: Benchmark and case study,”arXiv preprint arXiv:2409.08554, 2024

    2024 arXiv

  25. [33]

    Common 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, “Common V oice: A massively-multilingual speech corpus,” inProc. LREC, 2020, pp. 4218–4222

    2020

  26. [34]

    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. ICASSP, 2015, pp. 5206–5210

    2015

  27. [35]

    JSUT corpus: free large-scale Japanese speech corpus for end-to-end speech synthe- sis,

    R. Sonobe, S. Takamichi, and H. Saruwatari, “JSUT corpus: free large-scale Japanese speech corpus for end-to-end speech synthe- sis,”arXiv preprint arXiv:1711.00354, 2017

    2017 arXiv

  28. [36]

    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. ICML, 2021, pp. 5530–5540

    2021

  29. [37]

    CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit,

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit,” 2019

    2019

  30. [38]

    The Kaldi speech recognition toolkit,

    D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y . Qian, P. Schwarzet al., “The Kaldi speech recognition toolkit,” inProc. ASRU, 2011

    2011

  31. [39]

    with a rank of 16, applied to all linear modules. 4.6. Evaluation metrics We evaluated the results with WER for the English datasets and CER for the Japanese datasets. Additionally, we reported the recall and the precision of the rare words, following the method- ology describ...

  32. [40]

    Real- mednlp: Overview of real document-based medical natural lan- guage processing task,

    S. Yada, Y . Nakamura, S. Wakamiya, and E. Aramaki, “Real- mednlp: Overview of real document-based medical natural lan- guage processing task,” inProc. NTCIR-16, 2022, pp. 285–296

    2022

  33. [41]

    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

    2006 arXiv

  34. [42]

    Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,”Advances in neural information processing systems, vol. 33, pp. 17 022– 17 033, 2020

    2020

  35. [43]

    Contextualized streaming end-to-end speech recognition with trie-based deep bi- asing and shallow fusion,

    D. Le, M. Jain, G. Keren, S. Kim, Y . Shi, J. Mahadeokar, J. Chan, Y . Shangguan, C. Fuegen, O. Kalinliet al., “Contextualized streaming end-to-end speech recognition with trie-based deep bi- asing and shallow fusion,” inProc. Interspeech, 2021, pp. 1772– 1776

    2021

  36. [44]

    AI Speech: Azure AI Services,

    Microsoft, “AI Speech: Azure AI Services,” n.d. [On- line]. Available: https://azure.microsoft.com/en-us/products/ ai-services/ai-speech

  37. [45]

    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,” inInternational conference on machine learning, 2023, pp. 28 492–28 518

    2023

  38. [46]

    LoRA: Low-rank adaptation of large lan- guage models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large lan- guage models,” inProc. ICLR, 2022

    2022

  39. [47]

    Spell my name: Keyword boosted speech recognition,

    N. Jung, G. min Kim, and J. S. Chung, “Spell my name: Keyword boosted speech recognition,” 2021, pp. 6642–6646

    2021

  40. [48]

    Distribution of homonyms in Japanese text,

    M. Yamazaki, “Distribution of homonyms in Japanese text,” in Proc. LRW2021, 2021

    2021

Pith tools

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