Contextualized Automatic Speech Recognition with Dynamic Vocabulary Prediction and Activation

Kaixun Huang; Lei Xie; Linju Yang; Wei Ren; Zhennan Lin

REVIEW 4 major objections 5 minor 31 references

Contextualized Automatic Speech Recognition with Dynamic Vocabulary Prediction and Activation

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 Treating each contextual phrase as a single dynamic token, a frozen-encoder CTC model cuts rare-phrase word error by 72–76% and overall WER by 23–28%.

desk verdict Useful engineering extension of dynamic-vocabulary biasing to CTC, but the headline gains ride on an undisclosed confidence threshold and an uneven baseline comparison. read the letter →

arxiv 2505.23077 v1 pith:LHWUAMQS submitted 2025-05-29 cs.SD cs.CLeess.AS

Zhennan Lin , Kaixun Huang , Wei Ren , Linju Yang , Lei Xie This is my paper

classification cs.SDcs.CLeess.AS

keywords contextualbiasingdeepdynamicvocabularyCTCspeechrecognitionphrase-levelpredictionconfidence-activateddecodingWenet

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

This paper proposes a contextual biasing method for automatic speech recognition in which each phrase in a user-provided bias list is turned into a single dynamic token added to the CTC model's output vocabulary. The audio encoder stays frozen; only a bias module is trained, using a bias loss that aligns the phrase token with its frame-level occurrences. The paper reports that this approach reduces overall word error rate by 28.31% on LibriSpeech and 23.49% on WenetSpeech relative to a frozen-encoder baseline, while error on the biased phrases themselves falls by 72.04% and 75.69%, respectively. The method matters because it offers a way to specialize a recognizer to names and domain terms without retraining the full model.

What carries the argument

The key mechanism is the dynamic vocabulary of bias tokens, one per contextual phrase, appended to the CTC softmax output so that the model can emit a phrase as a single unit. A context encoder maps each phrase to an embedding, and a bias-aware module lets the audio hidden states attend to those embeddings to produce a contextualized representation. During decoding, the confidence-activated procedure scans the CTC peaks around a predicted bias token, computes the maximum posterior path probability as a confidence score, and replaces the last few predicted labels with the phrase only if that score exceeds a threshold scaled by the phrase length.

What would settle it

Sweep the confidence threshold on the LibriSpeech test-other set with bias list size 100 and record B-WER for each value; if B-WER at threshold zero or at a very high threshold is close to the reported 9.59, the confidence activation is not the source of the gain, and if the gain appears only in a narrow threshold range, the claimed robustness is falsified.

Watch

Extended reading notes

Core claim

The central claim is that phrase-level integrity, not just subword-level boosting, is what makes contextual biasing work, and that integrity can be achieved by predicting an explicit bias token for the whole phrase and then replacing the trailing subword labels with that phrase only when the CTC acoustic-path confidence is high. The paper demonstrates this by comparing two labeling strategies—word-by-word replacement and tail addition—and showing that tail addition, where the bias token is placed after the phrase's subword sequence, performs better, especially on longer Chinese phrases. The core discovery is that this combination of dynamic vocabulary prediction, a bias loss, and confidence-activated decoding produces large relative reductions in biased-phrase error while leaving non-contextual speech essentially unchanged.

Load-bearing premise

The reported gains depend on an unreported confidence threshold that determines when a predicted bias token triggers replacement of the preceding labels; if that threshold is not robust across phrase lengths, languages, or acoustic conditions, the headline improvements will not carry over to other settings.

Editorial extensions

If this is right

If the claim is right, encoder-based CTC models can perform contextual biasing without an autoregressive decoder, making the approach compatible with streaming and low-latency deployment.
The tail-addition labeling result implies that placing a whole-phrase token after the subword sequence is more robust for long phrases than replacing each subword, suggesting a design principle for future phrase-level losses.
The strong ablation for confidence activation implies that a model can safely enlarge its output vocabulary at inference time, provided the replacement decision is gated by an acoustic confidence check.
The near-neutral result with an empty bias list means a deployed system could add this bias module without regressing on ordinary speech, easing adoption.

Reading between the lines

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

A testable extension: the paper never reports the confidence threshold value or a sensitivity analysis, so a natural next experiment is to sweep that threshold on the LibriSpeech bias lists and map B-WER as a function of the threshold; a sharply peaked optimum would suggest the headline numbers are fragile, while a flat region would support robustness.
The same phrase-token idea could be transferred to RNN-T or attention decoders, but the CTC-peak-based confidence estimate would need a new alignment signal, so the transfer is not automatic.
The training procedure constructs the bias list by randomly sampling words from the transcript, so a realistic deployment with a much larger, noisier, or out-of-domain bias list may not exhibit the same 72% B-WER improvement; that is a concrete open question.

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.

Desk Editor's Note

read the letter

The paper does something real: it takes Sudo et al.'s dynamic-vocabulary idea from a decoder-based AED model and makes it work on a CTC encoder, which is a sensible scalability move. The TA labeling strategy, the bias loss, and the confidence-activated decoding are all reasonable engineering choices, and the internal ablations (Table 3) show each component pulls its weight. On both LibriSpeech and WenetSpeech, the method clearly improves contextual-phrase recognition over a frozen baseline. That is a genuine result worth reporting.

But the central numbers are more fragile than the abstract suggests. The confidence-activated decoding threshold is never disclosed. Table 3 shows that removing it takes WER from 6.81 to 10.31 and B-WER from 9.59 to 22.88, so this one hidden hyperparameter is responsible for most of the improvement. Without the threshold value, a sensitivity sweep, or at least a description of how it was tuned, the headline reductions cannot be reproduced or independently checked. That is a load-bearing omission, not a minor footnote.

The comparison to prior work is also uneven. Table 2 lists CPPN and DV with different baselines (8.88 and 5.98 WER, respectively) than the proposed DVPA-CTC (9.5). Relative B-WER improvement is a legitimate metric, but comparing percentage reductions across such different starting points is misleading. In absolute terms, DV's N=100 WER is 4.63 versus this paper's 6.81, so the claim of "outperforming" prior contextualized models is generous. The paper would be stronger if it retrained or at least reported the same baseline for all methods.

Smaller issues: the abstract and introduction give different relative reductions (28.31% vs 28.26%, 72.04% vs 71.16%), and no error bars or significance tests are reported anywhere. The threshold and the search window j are technically free parameters, but the bias-loss weights and sampling probability are also not swept.

For a reader working on contextual biasing for production CTC models, this paper is a useful data point. It deserves a serious referee and a round of revision: reveal the threshold, add a sensitivity analysis, and redo the comparison on a common baseline. The core idea is sound and the internal evidence is consistent, but the current presentation overstates the case.

Referee Report

4 major / 5 minor

Summary. The paper proposes an encoder-based CTC contextual ASR method that augments the vocabulary with phrase-level bias tokens. A context encoder produces embeddings for each phrase in the bias list, a bias-aware module fuses these with audio representations, and a bias loss trains the model to predict bias tokens. Two labeling strategies are compared (word replacement and tail addition), and a confidence-activated decoding post-processor replaces predicted subword sequences with the contextual phrase when a CTC-based confidence exceeds a threshold. Experiments on LibriSpeech and WenetSpeech report relative WER reductions of roughly 28% and 23% over a frozen baseline, and larger reductions in biased-word error rates. The paper also presents an ablation study and comparisons with CPPN and DV.

Significance. If the reported results are reproducible, the work is a useful incremental contribution: it adapts the phrase-level dynamic-vocabulary idea from AED systems to CTC encoders, and the confidence-activated decoding mechanism is a plausible way to control phrase insertion. The ablation study in Table 3 is informative because it isolates the contribution of each component. However, the strength of the claims is weakened by the undisclosed confidence threshold, the unmatched baselines used in the comparison to prior work, and the absence of error bars. These issues are fixable in revision, but they currently prevent the reader from verifying the central quantitative claims.

major comments (4)

[Section 2.3 and Table 3] The confidence threshold is never disclosed. Section 2.3 defines the activation rule using an unspecified scalar 'threshold' multiplied by phrase length ki, but the paper does not report its value, how it was initialized or tuned, or how sensitive the results are to it. This is load-bearing because Table 3 shows that removing confidence activation changes WER from 6.81 to 10.31 and B-WER from 9.59 to 22.88 on LibriSpeech N=100; the threshold is therefore responsible for most of the headline improvement. Please report the threshold value(s) used for LibriSpeech and WenetSpeech, the selection procedure (e.g., validation-set grid search), and a sensitivity sweep over threshold values. Without this, the central numbers cannot be independently checked.
[Table 2] The comparison to CPPN and DV is not apples-to-apples. CPPN's baseline has WER 8.88, DV's baseline has WER 5.98, and DVPA-CTC's baseline has WER 9.50; the relative B-WER improvements are computed from different base systems, so the claim that the proposed method 'outperforms' these approaches is unsupported. In fact, at N=100 DVPA-CTC has a higher absolute WER (6.81) than DV (4.63), while its relative B-WER reduction (-72.04%) is only slightly larger than DV's (-69.26%). Please either reimplement CPPN and DV with the same frozen base model or present a fully matched comparison, and report absolute and relative gains together with confidence intervals.
[Tables 2 and 4] The paper reports no error bars, standard deviations, or significance tests. All WER and B-WER figures are point estimates, presumably from a single run, and the relative improvements in the abstract and Section 3.3 are given to two decimal places (e.g., 28.32%). Given the sensitivity of the decoding post-processor to the threshold, please report results over at least three random seeds for the main configurations, or otherwise provide evidence that the observed differences are not within run-to-run variation.
[Abstract, Section 1, Section 3.3] The relative WER reduction on LibriSpeech is reported inconsistently: the abstract states 28.31%, the introduction states 28.26%, and Section 3.3 states 28.32%. Please align these numbers and state which configuration (e.g., N=100, test-other) each refers to. The contribution of this inconsistency is small, but it undermines confidence in the numerical care taken elsewhere.

minor comments (5)

[Equation (5)] The formula for Hdv appears dimensionally inconsistent as written: Linear(HCA) Linear(E^T)/sqrt(dk) does not obviously produce an n x T score matrix. Please clarify the notation and the intended shapes of the intermediate quantities.
[Section 3.4] The text 'Compared with the W A strategy' should read 'Compared with the WR strategy.'
[Section 2.1] There is a typo: 'the models ability' should be 'the model's ability.'
[Throughout] The names 'Librispeech' and 'WenetSpeech' are spelled inconsistently (LibriSpeech/Librispeech, WenetSpeech/WenetSpeech). Please unify them.
[Table 1] The table header says 'WER(U-WER/B-WER)' but the columns are only 'Strategy', 'Librispeech', 'WenetSpeech'; please clarify in the header or caption that the parenthesized values are the U-WER/B-WER breakdown.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported WER/B-WER gains are measured empirical results, not derived from the system's inputs; minor non-circular concerns are an undisclosed decoding threshold and post-hoc strategy selection.

full rationale

The paper's central claims are empirical: the relative WER reductions (28.31%, 23.49%) and contextual-phrase WER reductions (72.04%, 75.69%) are arithmetic transformations of measured WER and B-WER values in Tables 2 and 4, not quantities derived from the bias list, labels, or model equations. The bias list construction (e.g., from LibriSpeech bias lists in [20], or HanLP-extracted entities from WenetSpeech labels) is standard contextual-biasing evaluation practice, not a circular feed of the answer into the predictor: the model must still map audio to the correct phrase among distractors. The confidence-activated decoding rule (Sec. 2.3) is an algorithmic post-processor that replaces tokens with a bias-list phrase only when a path confidence exceeds a threshold; this is the intended mechanism of contextual biasing, and the model must correctly predict the bias token and the preceding subwords for the replacement to fire. The ablation in Table 3 shows the rule's contribution, but that is a system component, not a fitted parameter masquerading as a prediction. Self-citations ([19], [24], [25]) are prior work by the same group and standard tools/datasets; they are used for baseline comparison and experimental setup, not as load-bearing justification for the claimed improvement. The main legitimate concerns are reproducibility (the numeric threshold in Sec. 2.3 is never reported, and no sensitivity analysis is given) and that the TA strategy was chosen after seeing results on both test sets, which could inflate reported gains; however, these are methodological transparency issues, not circular derivations. No step in the paper reduces, by definition or by self-citation chain, to its own inputs, so the circularity score is minimal.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method depends on a handful of hyperparameters (loss weights, confidence threshold, search window, and bias-list sampling). The confidence threshold is the most consequential undisclosed parameter. The main invented artifact is the bias token, a modeling construct with no external evidence.

free parameters (4)

Loss weights lambda1, lambda2 = 0.3, 0.05
Section 2.1: hyperparameters in the joint loss L_total = lambda1*L_ctc + lambda2*L_bias; chosen by hand/validation.
Confidence threshold scalar = not reported
Section 2.3: threshold is scaled by phrase length ki; the base threshold value is not stated and is likely tuned on the validation set.
Search window range j = ki-2 to ki+2
Section 2.3: range for the starting frame when searching the best confidence path; chosen by hand.
Training bias-list sampling probability and count = 0.8, 2-10 words
Section 3.1: for each batch, with 80% probability 2 to 10 words are sampled as contextual phrases; these are hyperparameters.

assumptions (4)

standard math CTC formulation and standard ASR training assumptions
Invoked throughout; the paper builds on connectionist temporal classification as the base model (Section 2).
ad hoc to paper The first token of the context-encoder output represents each contextual phrase
Section 2.1: 'extract the vector at the first position of the label in the embedding sequence of each contextual phrase as the embedding vector ei.' This is a modeling choice without independent justification.
domain assumption Averaged multi-head attention scores are a sufficient confidence measure
Section 2.1 and 2.3 rely on attention scores and CTC posterior paths for confidence; the reliability of this measure is assumed.
domain assumption Bias lists are available at test time and independent of the audio
Deep biasing setup: the bias list is an input to the model (Section 2.1); no evidence that the method degrades gracefully with noisy or mismatched lists beyond the reported distractors.

invented entities (1)

Bias token <b_i>
purpose: Acts as a single output symbol representing an entire contextual phrase, appended after the phrase in the tail-addition strategy.
A new output vocabulary element introduced for phrase-level modeling; its benefit is demonstrated only in this paper's experiments, and it is a model construct rather than an independently observable quantity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Contextualized Automatic Speech Recognition with Dynamic Vocabulary Prediction and Activation." pith.science (2026). https://pith.science/paper/LHWUAMQS

@misc{pith2026250523077,
  author       = {Pith},
  title        = {Pith review of: Contextualized Automatic Speech Recognition with Dynamic Vocabulary Prediction and Activation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LHWUAMQS}},
  note         = {Machine review of arXiv:2505.23077}
}

read the original abstract

Deep biasing improves automatic speech recognition (ASR) performance by incorporating contextual phrases. However, most existing methods enhance subwords in a contextual phrase as independent units, potentially compromising contextual phrase integrity, leading to accuracy reduction. In this paper, we propose an encoder-based phrase-level contextualized ASR method that leverages dynamic vocabulary prediction and activation. We introduce architectural optimizations and integrate a bias loss to extend phrase-level predictions based on frame-level outputs. We also introduce a confidence-activated decoding method that ensures the complete output of contextual phrases while suppressing incorrect bias. Experiments on Librispeech and Wenetspeech datasets demonstrate that our approach achieves relative WER reductions of 28.31% and 23.49% compared to baseline, with the WER on contextual phrases decreasing relatively by 72.04% and 75.69%.

Figures

Figures reproduced from arXiv: 2505.23077 by the authors.

**Figure 1.** The overall architecture of the proposed method. 2.1. Model structure As shown in Fig.1, to ensure the scalability of the model, we exclusively incorporate the bias module into the encoder. Building upon the CTC model architecture, we introduced a context encoder, a bias-aware module, a multi-head attention-based output layer, and a bias projection layer. The audio encoder transforms the input audio features X int… view at source ↗

**Figure 2.** The confidence-activated decoding method. 2.3. confidence activation Our proposed confidence-activated decoding method is illustrated in Fig.2. As shown in Fig.2(a), when the model predicts the sequence y ′ = [. . . , A, lex, ander, < bi >, . . . ], we identify the termination frame pi as the CTC peak corresponding to the label < bi > and determine the starting frame pi−j as the j-th CTC peak before < bi >. Within… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

[1]

Alexander

Introduction In recent years, driven by advances in neural networks, end- to-end automatic speech recognition (E2E-ASR) has made re- markable progress [1, 2, 3, 4, 5, 6]. However, E2E-ASR mod- els depend heavily on their training data, resulting in a sig- nificant drop in recognition accuracy when encountering rare phrases (e.g., entity names and technica...

work page
[2]

Method This section introduces the proposed method based on the CTC architecture. To preserve the original recognition accuracy and facilitate the convergence of the bias module, we integrate the bias module into the pre-trained ASR model while keeping the original model parameters fixed, training only those associated with the bias module. arXiv:2505.230...

work page Pith review arXiv 2025
[3]

Experimental setup We train CTC/AED models as the baseline and the pre-trained ASR model, using the Wenet toolkit [25]

Experiments 3.1. Experimental setup We train CTC/AED models as the baseline and the pre-trained ASR model, using the Wenet toolkit [25]. The model takes an 80-dimensional Mel filterbank as input, with a frameshift of 10 ms and a frame length of 25 ms. SpecAugment [26] is ap- plied to enhance robustness. The audio encoder is a 12-layer Conformer, featuring...

work page
[4]

We investigate two labeling strategies to adaptively propagate phrase-level labels to frame-level out- puts

Conclusion In this paper, we propose a contextual deep biasing approach for speech recognition that leverages dynamic vocabulary pre- diction and activation. We investigate two labeling strategies to adaptively propagate phrase-level labels to frame-level out- puts. To ensure the integrity of subword sequences while miti- gating excessive boosting of cont...

work page
[5]

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,” in Proc. ICML , 2006, pp. 369–376

work page 2006
[6]

Sequence transduction with recurrent neural net- works,

A. Graves, “Sequence transduction with recurrent neural net- works,” in Proc. ICML, 2012

work page 2012
[7]

Attention-based models for speech recognition,

J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y . Ben- gio, “Attention-based models for speech recognition,” Advances in neural information processing systems, 2015

work page 2015
[8]

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 Proc. Interspeech, 2020, pp. 5036–5040

2020

Show all 31 references

[9]

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

W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” in Proc. ICASSP. IEEE, 2016, pp. 4960– 4964

2016
[10]

Learning phrase representations using rnn encoderdecoder for statistical machine translation,

K. Cho, B. van Merrienboer, . G¨ulehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoderdecoder for statistical machine translation,” in Proc. EMNLP, 2014

2014
[11]

Contextual speech recognition in end-to-end neural net- work systems using beam search,

I. Williams, A. Kannan, P. S. Aleksic, D. Rybach, and T. N. Sainath, “Contextual speech recognition in end-to-end neural net- work systems using beam search,” inProc. Interspeech, 2018, pp. 2227–2231

2018
[12]

End- to-end contextual speech recognition using class language models and a token passing decoder,

Z. Chen, M. Jain, Y . Wang, M. L. Seltzer, and C. Fuegen, “End- to-end contextual speech recognition using class language models and a token passing decoder,” inProc. ICASSP. IEEE, 2019, pp. 6186–6190

2019
[13]

Shallow-fusion end-to-end contextual biasing,

D. Zhao, T. N. Sainath, D. Rybach, P. Rondon, D. Bhatia, B. Li, and R. Pang, “Shallow-fusion end-to-end contextual biasing,” in Proc. Interspeech, 2019, pp. 1418–1422

2019
[14]

Class lm and word mapping for contextual biasing in end-to-end asr,

R. Huang, O. Abdel-hamid, X. Li, and G. Evermann, “Class lm and word mapping for contextual biasing in end-to-end asr,” in Proc. Interspeech, 2020, pp. 4348–4351

2020
[15]

Improved neural language model fu- sion for streaming recurrent neural network transducer,

S. Kim, Y . Shangguan, J. Mahadeokar, A. Bruguier, C. Fuegen, M. L. Seltzer, and D. Le, “Improved neural language model fu- sion for streaming recurrent neural network transducer,” in Proc. ICASSP. IEEE, 2021, pp. 7333–7337

2021
[16]

Joint grapheme and phoneme embeddings for contextual end-to-end asr,

Z. Chen, M. Jain, Y . Wang, M. L. Seltzer, and C. Fuegen, “Joint grapheme and phoneme embeddings for contextual end-to-end asr,” in Proc. Interspeech, 2019, pp. 3490–3494

2019
[17]

Cif-based collaborative decoding for end-to-end contextual speech recognition,

M. Han, L. Dong, S. Zhou, and B. Xu, “Cif-based collaborative decoding for end-to-end contextual speech recognition,” in Proc. ICASSP. IEEE, 2021, pp. 6528–6532

2021
[18]

Tree-constrained pointer generator for end-to-end contextual speech recognition,

G. Sun, C. Zhang, and P. C. Woodland, “Tree-constrained pointer generator for end-to-end contextual speech recognition,” in Proc. ASRU. IEEE, 2021, pp. 780–787

2021
[19]

Personalization of ctc speech recognition models,

S. Dingliwal, M. Sunkara, S. Ronanki, J. Farris, K. Kirchhoff, and S. Bodapati, “Personalization of ctc speech recognition models,” in Proc. SLT. IEEE, 2023, pp. 302–309

2023
[20]

Improving end-to-end contextual speech recognition with fine-grained contextual knowledge selection,

M. Han, L. Dong, Z. Liang, M. Cai, S. Zhou, Z. Ma, and B. Xu, “Improving end-to-end contextual speech recognition with fine-grained contextual knowledge selection,” in Proc. ICASSP. IEEE, 2022, pp. 8532–8536

2022
[21]

Deep context: end-to-end contextual speech recogni- tion,

G. Pundak, T. N. Sainath, R. Prabhavalkar, A. Kannan, and D. Zhao, “Deep context: end-to-end contextual speech recogni- tion,” in Proc. SLT. IEEE, 2018, pp. 418–425

2018
[22]

Context-aware transformer trans- ducer for speech recognition,

F.-J. Chang, J. Liu, M. Radfar, A. Mouchtaris, M. Omologo, A. Rastrow, and S. Kunzmann, “Context-aware transformer trans- ducer for speech recognition,” in Proc. ASRU. IEEE, 2021, pp. 503–510

2021
[23]

Contextualized end-to-end speech recognition with contextual phrase prediction network,

K. Huang, A. Zhang, Z. Yang, P. Guo, B. Mu, T. Xu, and L. Xie, “Contextualized end-to-end speech recognition with contextual phrase prediction network,” inProc. Interspeech, 2023, pp. 4933– 4937

2023
[24]

Contextualized 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. Kalinli, Y . Saraf, and M. L. Seltzer, “Contextualized streaming end-to-end speech recognition with trie-based deep biasing and shallow fusion,” in Proc. Interspeech, 2021, pp. 1772–1776

2021
[25]

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,” in Proc. ASRU. IEEE, 2023, pp. 1–8

2023
[26]

Contextualized automatic speech recognition with dynamic vo- cabulary,

Y . Sudo, Y . Fukumoto, M. Shakeel, Y . Peng, and S. Watanabe, “Contextualized automatic speech recognition with dynamic vo- cabulary,” inProc. SLT. IEEE, 2024, pp. 78–85

2024
[27]

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,” in Proc. ICASSP. IEEE, 2015, pp. 5206–5210

2015
[28]

Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

B. Zhang, H. Lv, P. Guo, Q. Shao, C. Yang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zeng et al., “Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” in Proc. ICASSP. IEEE, 2022, pp. 6182–6186

2022
[29]

Wenet: Production oriented stream- ing and non-streaming end-to-end speech recognition toolkit,

Z. Yao, D. Wu, X. Wang, B. Zhang, F. Yu, C. Yang, Z. Peng, X. Chen, L. Xie, and X. Lei, “Wenet: Production oriented stream- ing and non-streaming end-to-end speech recognition toolkit,” in Proc. Interspeech, 2021, pp. 1–5

2021
[30]

Specaugment: A simple data augmentation method for automatic speech recognition,

D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” inProc. Interspeech, 2019, pp. 2613–2617

2019
[31]

Cb-conformer: Contextual biasing conformer for biased word recognition,

Y . Xu, B. Liu, Q. Huang, X. Song, Z. Wu, S. Kang, and H. Meng, “Cb-conformer: Contextual biasing conformer for biased word recognition,” in Proc. ICASSP. IEEE, 2023, pp. 1–5

2023

Pith tools

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

[1] [1]

Alexander

Introduction In recent years, driven by advances in neural networks, end- to-end automatic speech recognition (E2E-ASR) has made re- markable progress [1, 2, 3, 4, 5, 6]. However, E2E-ASR mod- els depend heavily on their training data, resulting in a sig- nificant drop in recognition accuracy when encountering rare phrases (e.g., entity names and technica...

work page

[2] [2]

Method This section introduces the proposed method based on the CTC architecture. To preserve the original recognition accuracy and facilitate the convergence of the bias module, we integrate the bias module into the pre-trained ASR model while keeping the original model parameters fixed, training only those associated with the bias module. arXiv:2505.230...

work page Pith review arXiv 2025

[3] [3]

Experimental setup We train CTC/AED models as the baseline and the pre-trained ASR model, using the Wenet toolkit [25]

Experiments 3.1. Experimental setup We train CTC/AED models as the baseline and the pre-trained ASR model, using the Wenet toolkit [25]. The model takes an 80-dimensional Mel filterbank as input, with a frameshift of 10 ms and a frame length of 25 ms. SpecAugment [26] is ap- plied to enhance robustness. The audio encoder is a 12-layer Conformer, featuring...

work page

[4] [4]

We investigate two labeling strategies to adaptively propagate phrase-level labels to frame-level out- puts

Conclusion In this paper, we propose a contextual deep biasing approach for speech recognition that leverages dynamic vocabulary pre- diction and activation. We investigate two labeling strategies to adaptively propagate phrase-level labels to frame-level out- puts. To ensure the integrity of subword sequences while miti- gating excessive boosting of cont...

work page

[5] [5]

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,” in Proc. ICML , 2006, pp. 369–376

work page 2006

[6] [6]

Sequence transduction with recurrent neural net- works,

A. Graves, “Sequence transduction with recurrent neural net- works,” in Proc. ICML, 2012

work page 2012

[7] [7]

Attention-based models for speech recognition,

J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y . Ben- gio, “Attention-based models for speech recognition,” Advances in neural information processing systems, 2015

work page 2015

[8] [8]

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 Proc. Interspeech, 2020, pp. 5036–5040

2020