Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM

T0 review · 5 major / 6 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 A retrieval front-end for speech LLMs lets contextual ASR scale from dozens to 200,000 bias words without fine-tuning the ASR model.

desk verdict A plausible retrieval-based biasing framework with a real, fixable flaw in Eq. 6 and overstated headline numbers; worth reviewing, needs major revision. read the letter →

arxiv 2505.19179 v1 pith:PRUPLA4M submitted 2025-05-25 cs.SD eess.ASeess.SP

Xun Gong , Anqi Lv , Zhiming Wang , Huijia Zhu , Yanmin Qian This is my paper
classification cs.SDeess.ASeess.SP
keywords contextualbiasingspeechlargelanguagemodelbiasretrievalcontrastivelearninghomophonecurriculumrarewordrecognitionscalableASRretrieval-augmentedgeneration
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 BR-ASR, a retrieval front-end for contextual biasing in speech large language models. Instead of stuffing every bias word into the model's prompt, which causes hallucinations and fails beyond dozens of entries, it learns a contrastive embedding space that matches an utterance against a large database of bias words and returns only a short candidate list. The paper claims this keeps performance stable up to 200,000 bias entries, cuts biased-word error to 2.8%/7.1% on LibriSpeech test-clean/-other with 2,000 bias words, and adds only about 20 ms per query. A second mechanism, a phoneme-aware curriculum that increases homophone negatives during training, suppresses errors where the retrieved word sounds like another word. If the claims hold, the bottleneck is no longer the size of the bias list but the quality of the retrieved candidates.

What carries the argument

The load-bearing mechanism is a contrastive speech-and-bias embedding space trained with a CLAP-style objective: a speech encoder produces frame-level features, a pooling network (attention pooling is best) collapses them to a vector, and a bias encoder maps each bias word, either through synthesized audio in the AcousticBias variant or through the LLM text encoder in the TextualBias variant, to a vector of the same dimension. The dot product between the pooled utterance and candidate bias vectors is scored, temperature-scaled, and optimized with a symmetric contrastive loss over hard negatives. The second mechanism is the dynamic curriculum: the paper builds homophone sets using grapheme-to-phoneme conversion and Levenshtein distance on phoneme sequences, then ramps the sampling ratio of homophone negatives with a sigmoidal schedule while adding an embedding-dispersion regularizer. Together these mechanisms let the framework retrieve the correct word at 93% accuracy while discarding 99.99% of the database, and keep top-50 homophone recall low enough that the downstream ASR is not misled.

What would settle it

Take a held-out audio set with known rare-word transcripts and a bias database of 200,000 entries deliberately crowded with near-homophones of the target words, then measure the fraction of sentences where the oracle bias word is in the top-50 retrieved candidates; if that recall falls below roughly 99%, B-WER will degrade by more than the reported 2.9 absolute points at 200k, because retrieval errors cannot be corrected downstream.

Watch

Extended reading notes

Core claim

The central discovery is that scaling contextual ASR can be recast as a retrieval problem rather than a prompt-engineering or decoder-bias problem. With frozen speech and bias encoders taken from the host SpeechLLM, the paper trains only a pooling network and a contrastive head so that a pooled utterance vector and a bias-word vector have high dot-product similarity when the word is the correct rare term. Retrieval then prunes 99.99% of a 200k-entry database, from 200,000 candidates to a top-50 list, and the ASR model sees only those candidates. On the LibriSpeech Rare5k benchmark, this yields B-WER of 2.8%/7.1% at N=2000, a 45% relative improvement over prior approaches, while WER/B-WER degrade only 0.3/2.9 absolute points when the database grows from 2k to 200k entries. The paper further shows the framework is backend-agnostic, improving both prompt-based and bias-encoder SpeechLLMs and a third-party SLAM-ASR system without modifying the ASR model.

Load-bearing premise

The frozen pretrained encoders, trained only through a pooling network and a contrastive head, put the correct bias word in the top-50 retrieved candidates across a 200,000-entry bias list; if that retrieval recall breaks, the ASR cannot recover the word.

Editorial extensions

If this is right

  • A bias list of 200,000 entries becomes usable in a SpeechLLM without fine-tuning the ASR model, with only 0.3/2.9 absolute WER/B-WER degradation on test-other.
  • Retrieval can be decoupled from decoding, so the same trained matcher can be plugged into different ASR backends, including prompt-based and bias-encoder SpeechLLMs and a CTC-filter SLAM-ASR system.
  • Because only 50 candidates reach the ASR, prompt length stays short even when the database is huge, avoiding the hallucination and context-limit failures of full bias-list prompting.
  • The homophone curriculum reduces retrieval-level confusion, which is the main source of B-WER degradation at N=2000, so further bias-list growth is limited by retrieval recall rather than decoder capacity.

Reading between the lines

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

  • An editor-level extrapolation: the same retrieval front-end should transfer to other conditional speech tasks with large vocabularies, such as speech translation with glossary constraints, because the contrastive objective is not ASR-specific.
  • The paper measures recall on LibriSpeech, where rare words are mostly book-derived names; the authors do not test dynamically added user-specific terms, so online insertion of new bias words into the vector index and OOV embeddings is the next practical bottleneck.
  • A direct test of the framework's claim would be to run it on accented or noisy speech: frozen encoder priors may not keep acoustic variants of a target word near its canonical embedding, and retrieval recall would drop before the ASR can intervene.
  • The 20ms latency number assumes a fixed 4096-dimensional vector and a static GPU index; streaming deployment would require deciding when to run retrieval, per utterance, per chunk, or on a detected rare-word trigger, which the paper leaves open.
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

5 major / 6 minor

Summary. The paper proposes BR-ASR, a retrieval-based contextual biasing framework for SpeechLLM-based ASR. The method freezes the host system's speech and bias encoders, trains only a speech pooling network and a contrastive head, and retrieves a small set of candidate bias words from a large database via FAISS before feeding them to the SpeechLLM. A second claimed contribution is a phoneme-aware dynamic curriculum that is intended to suppress homophone confusion during contrastive training. Experiments on LibriSpeech test-clean/-other with the Rare5k bias list report B-WER of 2.8%/7.1% at 2000 bias words, scalability to 200k database entries with 0.3%/2.9% absolute WER/B-WER degradation, and 20 ms GPU retrieval latency.

Significance. If the claims hold, the paper addresses a practically important bottleneck: contextual biasing at bias-list scales well beyond the dozens of entries that prompt-based SpeechLLM methods can handle. The decoupling of retrieval from ASR decoding is a sensible design, and the framework's ability to plug into both Qwen-Audio and SLAM-ASR backends is a useful generalization result. The ablation matrix in Table 2 is informative about pooling and bias-encoding choices, and the FAISS latency numbers give a concrete efficiency check. However, the paper's central mechanistic claim for dynamic curriculum learning is not substantiated as written: under the stated frozen-encoder prior, the regularization term in Eq. (6) is a constant and cannot provide a training signal. Several headline quantitative claims are also inconsistent with the presented tables. These issues are load-bearing for the paper's two claimed innovations and for its SOTA/scalability statements, even though the underlying retrieval idea and the B-WER numbers are not directly falsified by the review.

major comments (5)
  1. [Section 2.2.2 and Eq. (6)] The homophone regularization term is a constant under the stated frozen-encoder design, so the dynamic-curriculum ablations cannot be attributed to the stated objective. Section 2.2.2 states that E_s and E_b are fully frozen, and Eq. (2) defines bias features as output of E_b. Therefore, in Eq. (6), both b_i and h_j are fixed vectors, and the term λ Σ ||b_i - h_j||^2 contributes no gradient to the only trainable parameters, namely the speech Pool and the contrastive head. The RecallH#50 and B-WER changes in Table 2 must instead come from the hard-negative sampling schedule, but Eq. (5) never states how α_n modifies the sampling distribution, and the 'off' configuration in Table 2 is not defined. Please either make the bias embeddings trainable, remove Eq. (6), or specify the exact sampling mechanism and ablate it directly.
  2. [Abstract, Section 3.2, Table 1] The headline relative improvements do not match the table values. The abstract claims a 45% relative improvement over prior methods at N=2000, and Section 3.2 claims a 17.1% relative improvement over USTR-CT's best baseline at N=2000. From Table 1, Prompt-QwenAudio+Bias Retrieval at N=2000 gives B-WER 2.8/7.1 versus USTR-CT's 3.0/7.6, which is only about 6.6% relative improvement on either test set; the 17.1% figure appears nowhere in the table. Please recompute these percentages and state the exact baseline and N used for each claimed improvement.
  3. [Abstract, Section 3.4, Figure 2] The claimed 0.3% absolute WER degradation at 200k entries is not supported by the reported data. Figure 2 plots only B-WER, and Section 3.4 reports only that B-WER degrades by 2.9% (from 9.9% to 12.8%) for the BiasEncoder method when |B_total| grows from 2k to 200k. No table or figure reports WER as a function of database size. Please provide the WER measurements or remove the 0.3% WER claim.
  4. [Section 3.1.1, Abstract] The retrieval metrics are incompletely specified and the abstract's retrieval accuracy numbers are unverifiable. RecallB and RecallH are written as ratios of sums with ambiguous notation, and the footnote defining Recall@X / Recall#X is unclear about whether the reported numbers are percentages or counts. The abstract's '93% / 91% retrieval accuracy' is not defined anywhere and does not appear in Table 2 or any other table. Please define the metrics precisely, clarify the direction of RecallH (the arrow suggests lower is better, which is unusual for a recall), and report all claimed retrieval accuracy numbers in a table.
  5. [Section 3.2, Table 1] The two closest large-scale baselines mentioned in the introduction, Dual-NAM [20] and VQ-RAG [21], are not included in the experimental comparison. Since the paper's scalability claim is explicitly 'where traditional methods generally fail', omitting methods that operate at 40k–100k bias entries leaves the new-SOTA claim unsupported. Please add these comparisons, or explain concretely why they cannot be run on the LibriSpeech/Rare5k setup.
minor comments (6)
  1. [Abstract] There are typographical errors that should be corrected: 'The is a general framework' should read 'This is a general framework'; 'unrealiable' should be 'unreliable'; 'fine-graind' should be 'fine-grained'; 'withou' should be 'without'; and 'Retreival' should be 'Retrieval'.
  2. [Section 2.2.3] The sentence 'we use grapheme-to-phoneme (G2P) to generate phoneme sequence and and construct homophone set' contains a doubled 'and' and should be reworded.
  3. [Table 1] The 'X' entries in Table 1 are not explained in the caption; please state that they denote catastrophic failure or context-length overflow, as described in the text.
  4. [Section 3.1.2] The phrase 'the unchat version' is informal; please clarify what this means (e.g., a base model without chat fine-tuning) and cite the exact model version.
  5. [Section 3.1.2] The 'retrieval threshold' values of 50, 50, and 10 are described with an undefined term; if this is the retrieval top-K, please name it consistently and explain why different thresholds are used for different backends.
  6. [References] Reference [14] is a self-citation to 'Proc. Interspeech, 2025' with the same title as this manuscript; if this work is already accepted or under submission elsewhere, the overlap should be disclosed to the editor.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the retrieval results are evaluated against external baselines on held-out LibriSpeech, and the self-citations are not load-bearing; the inert Eq. 6 regularizer is a correctness concern, not a circular prediction.

full rationale

BR-ASR's central evaluation is empirically self-contained. The contrastive retriever is trained on LibriSpeech train-960h with oracle bias words as in-batch positives and evaluated on held-out test-clean/-other, which is a standard supervised setup rather than a circular reuse of test labels. The homophone sets in Eq. 4 are constructed from G2P phoneme distances, not from the test transcripts or from the B-WER targets, so the curriculum objective is not defined in terms of the evaluation quantity. External baselines (DB-NNLM, USTR-CT, CTC-Filter) are used for the reported SOTA comparison, so the central claim does not reduce to a self-citation chain. Self-citations exist ([11] provides the QwenAudio contextual baselines; [28-30] are cited for curriculum learning), but the claimed improvements are measured against these reproduced baselines and against independent prior systems, so the self-citations are not load-bearing. The one notable by-construction issue is Eq. 6: because E_s and E_b are fully frozen per Section 2.2.2, both b_i and h_j are constant outputs of the frozen bias encoder, making ||b_i - h_j||^2 a constant with respect to the only trainable components (Pool and the contrastive head). Thus the regularization term cannot by itself disperse homophones, and the DCL ablation gains in Table 2 must be attributed to the alpha_n hard-negative sampling schedule in Eq. 5 rather than to Eq. 6. This is a mechanistic/correctness gap in the paper's description, not a case where a predicted quantity is equivalent to a fitted input or to a self-cited theorem; the reported B-WER numbers are not shown to be forced by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework rests on frozen pretrained encoders and several hand-tuned hyperparameters (θ, α, λ, retrieval thresholds). It introduces no invented entities; the contrastive objective and G2P homophone mining are standard techniques applied to a new domain.

free parameters (5)
  • Homophone distance threshold θ = 2
    Chosen by hand to define homophone sets via Levenshtein distance on G2P sequences (Eq. 4).
  • Curriculum schedule α_min, α_max, γ = 0.01, 0.5, 0.05
    Hand-tuned sigmoidal ramp controlling homophone hard-negative sampling ratio (Eq. 5).
  • Homophone dispersion weight λ = 0.1
    Hand-tuned weight for the homophone dispersion regularization term (Eq. 6).
  • Retrieval top-K threshold = 50, 50, 10
    Number of retrieved bias candidates fed to Prompt-QwenAudio, BiasEncoder-QwenAudio, and Prompt-SLAM-ASR respectively; directly affects recall ceiling and prompt size.
  • Contrastive temperature τ = init 0.007, clipped at 100
    Standard CLAP temperature controlling the sharpness of the contrastive distribution (Eq. 3).
assumptions (4)
  • domain assumption Frozen pretrained encoders (speech encoder and LLM text encoder from Qwen-Audio/SLAM-ASR) provide a representation space where cross-modal dot-product retrieval can work after a light trainable pooling layer.
    Invoked in Section 2.2.2 'Frozen Encoding Priors' and in Eq. (2)-(3); the encoders are not fine-tuned, so retrieval quality depends entirely on their existing alignment.
  • domain assumption Grapheme-to-phoneme conversion is accurate for rare OOV words in the bias database.
    Used in Eq. (4) to construct homophone sets; errors in G2P would create incorrect hard negatives and degrade the curriculum.
  • domain assumption The oracle bias list B_sentence from Rare5k is correct and complete for every LibriSpeech sentence.
    Used as ground-truth positives in the contrastive training objective (Section 2.2.1); incomplete or wrong oracle lists would train false positives/negatives.
  • domain assumption LibriSpeech and the Rare5k rare-word set are representative of real-world contextual biasing conditions.
    All experiments and database scaling tests use this single corpus; generalization to other domains or accents is not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM." pith.science (2026). https://pith.science/paper/PRUPLA4M

@misc{pith2026250519179,
  author       = {Pith},
  title        = {Pith review of: BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRUPLA4M}},
  note         = {Machine review of arXiv:2505.19179}
}
read the original abstract

While speech large language models (SpeechLLMs) have advanced standard automatic speech recognition (ASR), contextual biasing for named entities and rare words remains challenging, especially at scale. To address this, we propose BR-ASR: a Bias Retrieval framework for large-scale contextual biasing (up to 200k entries) via two innovations: (1) speech-and-bias contrastive learning to retrieve semantically relevant candidates; (2) dynamic curriculum learning that mitigates homophone confusion which negatively impacts the final performance. The is a general framework that allows seamless integration of the retrieved candidates into diverse ASR systems without fine-tuning. Experiments on LibriSpeech test-clean/-other achieve state-of-the-art (SOTA) biased word error rates (B-WER) of 2.8%/7.1% with 2000 bias words, delivering 45% relative improvement over prior methods. BR-ASR also demonstrates high scalability: when expanding the bias list to 200k where traditional methods generally fail, it induces only 0.3 / 2.9% absolute WER / B-WER degradation with a 99.99% pruning rate and only 20ms latency per query on test-other.

Figures

Figures reproduced from arXiv: 2505.19179 by the authors.

Figure 1
Figure 1. The proposed Bias Retrieval Framework with training (1) and decoding (2) procedures for Contextual SpeechLLM. the test-other subset, resolving the long-tail coverage issues. With the help of FAISS [23], the retrieval latency is only 20ms when querying 200k entries on GPU. Crucially, our framework demonstrates strong generalization: when integrated with third￾party ASR systems, it consistently improves rare-word reco… view at source ↗
Figure 2
Figure 2. The B-WER (%) performance across different bias database (Btotal) scales from 10 to 200k entries. Specially, S = |Bsentence|). |Bsentence| (In [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System

    cs.CL 2025-08 conditional novelty 6.0 of 10

    A cross-attention term retriever estimates which terminology appears in speech and, when its top-k terms are added to the prompt, improves SLM terminology accuracy by 6-17%.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [20]

    Retrieval-augmented gen- eration for large language models: A survey,

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, M. Wang, and H. Wang, “Retrieval-augmented gen- eration for large language models: A survey,”arXiv preprint arXiv:2312.10997, 2023

    arXiv 2023

  2. [21]

    DOC-RAG: ASR language model personaliza- tion with domain-distributed co-occurrence retrieval augmenta- tion,

    P. Mathur, Z. Liu, K. Li, Y . Ma, G. Karen, Z. Ahmed, D. Manocha, and X. Zhang, “DOC-RAG: ASR language model personaliza- tion with domain-distributed co-occurrence retrieval augmenta- tion,” inProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evalua- tion (LREC-COLING 2024), 2024, pp. 5132–5139

    work page 2024

  3. [1]

    Introduction The integration of speech processing into large language mod- els (SpeechLLMs) has advanced speech tasks like automatic speech recognition (ASR), achieving state-of-the-art perfor- mance in general domains [1], making them more applicable in real-world scenarios such as voice assistants. However, chal- lenges arise in accurately recognizing a...

    arXiv 2025

  4. [2]

    Related bias words areb 1, b2,· · ·

    Methodology 2.1. Revisiting Contextual Biasing in SpeechLLM-based ASR Systems The contextual ASR (C-ASR) can be formally defined through three core components: (1) the input audio spectrogramX∈ RM×T , whereMdenotes mel-frequency bins andTtemporal frames, (2) a predefined bias listB={b 1, b2, . . . , bN }contain- ingNdomain-specific terms or rare entities ...

    work page

  5. [3]

    en-US-AnaNeural

    Experiments 3.1. Experimental Setup 3.1.1. Dataset and Metrics Experiments were conducted using LibriSpeech [31] for both training and evaluation, where bias words are marked by Rare5k [22]. Rare5k identifies the top 5k most frequent words as common, with the remaining 209.2k words classified as rare in the LibriSpeech train-960h set. Following the same s...

    work page

  6. [4]

    Conclusion In this paper, we present BR-ASR: a bias retrieval framework that fundamentally resolves the scalability bottleneck in contex- tual ASR. Our proposed method has 4 innovations: (1) Cross- modal contrastive learning aligns speech inputs with textual bias entries via pretrained speech and bias encoders, achiev- ing 93% retrieval accuracy while pru...

    work page 2000

  7. [5]

    Acknowledgements This work was supported in part by China NSFC projects under Grants 62122050 and 62071288, in part by Shanghai Munici- pal Science and Technology Commission Project under Grant 2021SHZDZX0102, and in part by Ant Group and Ant Group Research Intern Program

    work page

  8. [6]

    WavChat: A survey of spoken dialogue models,

    S. Ji, Y . Chen, M. Fang, J. Zuo, J. Lu, H. Wang, Z. Jiang, L. Zhou, S. Liu, X. Chenget al., “WavChat: A survey of spoken dialogue models,”arXiv preprint arXiv:2411.13577, 2024

    arXiv 2024

Show all 39 references
  1. [7]

    Deep context: End-to-end contextual speech recog- nition,

    G. Pundak, T. N. Sainath, R. Prabhavalkar, A. Kannan, and D. Zhao, “Deep context: End-to-end contextual speech recog- nition,” in2018 IEEE Spoken Language Technology Workshop (SLT), 2018, pp. 418–425

    2018

  2. [8]

    PromptASR for contextualized ASR with controllable style,

    X. Yang, W. Kang, Z. Yao, Y . Yang, L. Guo, F. Kuang, L. Lin, and D. Povey, “PromptASR for contextualized ASR with controllable style,”arXiv preprint arXiv:2309.07414, 2023

    2023 arXiv

  3. [9]

    Text only domain adaptation with phoneme guided data splicing for end- to-end speech recognition,

    W. Wang, X. Gong, H. Shao, D. Yang, and Y . Qian, “Text only domain adaptation with phoneme guided data splicing for end- to-end speech recognition,” inInterspeech 2023, 2023, pp. 3347– 3351

    2023

  4. [10]

    LaMP: When large language models meet personalization,

    A. Salemi, S. Mysore, M. Bendersky, and H. Zamani, “LaMP: When large language models meet personalization,”arXiv preprint arXiv:2304.11406, 2023

    2023 arXiv

  5. [11]

    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,” inIEEE International Conference on Acoustics, Speech and Signal P...

    2024

  6. [12]

    CTC-assisted LLM-based contextual ASR,

    G. Yang, Z. Ma, Z. Gao, S. Zhang, and X. Chen, “CTC-assisted LLM-based contextual ASR,” in2024 IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 126–131

    2024

  7. [13]

    Con- textual spelling correction with large language models,

    G. Song, Z. Wu, G. Pundak, A. Chandorkar, K. Joshi, X. Velez, D. Caseiro, B. Haynor, W. Wang, N. Siddharthaet al., “Con- textual spelling correction with large language models,” in2023 IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU), 2023, pp. 1–8

    2023

  8. [14]

    Contextual ASR error handling with LLMs augmentation for goal-oriented conversational AI,

    Y . Asano, S. Hassan, P. Sharma, A. Sicilia, K. Atwell, D. Litman, and M. Alikhani, “Contextual ASR error handling with LLMs augmentation for goal-oriented conversational AI,”arXiv preprint arXiv:2501.06129, 2025

    2025 arXiv

  9. [15]

    End-to-end speech recognition contextualization with large language models,

    E. Lakomkin, C. Wu, Y . Fathullah, O. Kalinli, M. L. Seltzer, and C. Fuegen, “End-to-end speech recognition contextualization with large language models,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 12 406–12 410

    2024

  10. [16]

    Contextual biasing speech recognition in speech-enhanced large language model,

    X. Gong, A. Lv, Z. Wang, and Y . Qian, “Contextual biasing speech recognition in speech-enhanced large language model,” Proc. Interspeech, pp. 257–261, 2024

    2024

  11. [17]

    Ranking and selec- tion of bias words for contextual bias speech recognition,

    H. Hou, X. Gong, W. Zhang, and Y . Qian, “Ranking and selec- tion of bias words for contextual bias speech recognition,”Proc. Interspeech, 2025

    2025

  12. [18]

    Prompting large language models for zero-shot domain adaptation in speech recognition,

    Y . Li, Y . Wu, J. Li, and S. Liu, “Prompting large language models for zero-shot domain adaptation in speech recognition,” in2023 IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU), 2023, pp. 1–8

    2023

  13. [19]

    Br-asr: Effi- cient and scalable bias retrieval framework for contextual biasing asr in speech llm,

    X. Gong, A. Lv, Z. Wang, Z. Huijia, and Y . Qian, “Br-asr: Effi- cient and scalable bias retrieval framework for contextual biasing asr in speech llm,”Proc. Interspeech, 2025

    2025

  14. [22]

    LA-RAG: Enhancing LLM-based ASR ac- curacy with retrieval-augmented generation,

    S. Li, H. Shang, D. Wei, J. Guo, Z. Li, X. He, M. Zhang, and H. Yang, “LA-RAG: Enhancing LLM-based ASR ac- curacy with retrieval-augmented generation,”arXiv preprint arXiv:2409.08597, 2024

    2024 arXiv

  15. [23]

    GEC-RAG: Improv- ing generative error correction via retrieval-augmented genera- tion for automatic speech recognition systems,

    A. Robatian, M. Hajipour, M. R. Peyghan, F. Rajabi, S. Amini, S. Ghaemmaghami, and I. Gholampour, “GEC-RAG: Improv- ing generative error correction via retrieval-augmented genera- tion for automatic speech recognition systems,”arXiv preprint arXiv:2501.10734, 2025

    2025 arXiv

  16. [24]

    Contextualization of ASR with LLM using phonetic retrieval-based augmentation,

    Z. Lei, X. Na, M. Xu, E. Pusateri, C. Van Gysel, Y . Zhang, S. Han, and Z. Huang, “Contextualization of ASR with LLM using phonetic retrieval-based augmentation,”arXiv preprint arXiv:2409.15353, 2024

    2024 arXiv

  17. [25]

    Optimizing large-scale context retrieval for end-to-end ASR,

    Z. Huang, D. Caseiro, K. Joshi, C. Li, P. Rondon, Z. Wu, P. Zadrazil, and L. Zhou, “Optimizing large-scale context retrieval for end-to-end ASR,” inProc. Interspeech, 2024, pp. 4573–4577

    2024

  18. [26]

    Optimizing contextual speech recognition us- ing vector quantization for efficient retrieval,

    N. Flemotomos, R. Hsiao, P. Swietojanski, T. Hori, D. Can, and X. Zhuang, “Optimizing contextual speech recognition us- ing vector quantization for efficient retrieval,”arXiv preprint arXiv:2411.00664, 2024

    2024 arXiv

  19. [27]

    Contextual- ized streaming end-to-end speech recognition with trie-based deep biasing and shallow fusion,

    D. Le, M. Jain, G. Keren, S. Kim, Y . Shi, J. Mahadeokar, J. Chan, Y . Shangguan, C. Fuegen, O. Kalinliet al., “Contextual- ized streaming end-to-end speech recognition with trie-based deep biasing and shallow fusion,”arXiv preprint arXiv:2104.02194, 2021

    2021 arXiv

  20. [28]

    Billion-scale similarity search with GPUs,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with GPUs,”IEEE Transactions on Big Data, vol. 7, no. 3, pp. 535–547, 2019

    2019

  21. [29]

    Longfnt: Long-form speech recognition with factorized neu- ral transducer,

    X. Gong, Y . Wu, J. Li, S. Liu, R. Zhao, X. Chen, and Y . Qian, “Longfnt: Long-form speech recognition with factorized neu- ral transducer,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  22. [30]

    Factor- ized aed: Factorized attention-based encoder-decoder for text- only domain adaptive asr,

    X. Gong, W. Wang, H. Shao, X. Chen, and Y . Qian, “Factor- ized aed: Factorized attention-based encoder-decoder for text- only domain adaptive asr,” inICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  23. [31]

    Advanced long-content speech recognition with factorized neu- ral transducer,

    X. Gong, Y . Wu, J. Li, S. Liu, R. Zhao, X. Chen, and Y . Qian, “Advanced long-content speech recognition with factorized neu- ral transducer,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1803–1815, 2024

    2024

  24. [32]

    CLAP learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “CLAP learning audio concepts from natural language supervision,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  25. [33]

    Knowledge transfer and distil- lation from autoregressive to non-autoregressive speech recogni- tion,

    X. Gong, Z. Zhou, and Y . Qian, “Knowledge transfer and distil- lation from autoregressive to non-autoregressive speech recogni- tion,” inProc. Interspeech 2022, 2022, pp. 2618–2622

    2022

  26. [34]

    Speaker embedding augmentation with noise distribution match- ing,

    X. Gong, Z. Chen, Y . Yang, S. Wang, L. Wang, and Y . Qian, “Speaker embedding augmentation with noise distribution match- ing,” in2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2021, pp. 1–5

    2021

  27. [35]

    Layer-wise fast adaptation for end-to-end multi-accent speech recognition,

    Y . Qian, X. Gong, and H. Huang, “Layer-wise fast adaptation for end-to-end multi-accent speech recognition,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 30, p. DOI:10.1109/TASLP.2022.3198546, 2022

    2022

  28. [36]

    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,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), 2015, pp. 5206–5210

    2015

  29. [37]

    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

  30. [38]

    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. Zhang, and X. Chen, “An embarrassingly simple approach for LLM with strong ASR capacity,”arXiv preprint arXiv:2402.08846, 2024

    2024 arXiv

  31. [39]

    Improving large-scale deep biasing with phoneme features and text-only data in streaming transducer,

    J. Qiu, L. Huang, B. Li, J. Zhang, L. Lu, and Z. Ma, “Improving large-scale deep biasing with phoneme features and text-only data in streaming transducer,” in2023 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU), 2023, pp. 1–8

    2023

Pith tools

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