Pith. sign in

REVIEW 4 major objections 6 minor 40 references

K-step prediction — a decoder branch that forecasts future subword tokens — lets trie-based biasing reward rare-word prefixes accurately, removing the need for reward revocation and cutting WER on a Singapore English test from 30.86% to 12.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A future-token prediction branch in Whisper gates trie-based biasing rewards, letting greedy decoding recognize rare words without a beam-search reward revocation step.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection K-step gating is a legitimate new twist on trie biasing, but the paper doesn't yet show the gate works where it must, and the headline gain mostly comes from fine-tuning and plain trie. the 4 major comments →

arxiv 2509.09196 v1 pith:6VBI2BZ4 submitted 2025-09-11 cs.CL cs.AI

Efficient Trie-based Biasing using K-step Prediction for Rare Word Recognition

Chin Yuen Kwok , Jia Qi yip This is my paper
classification cs.CL cs.AI
keywords contextual biasingtrie-based biasingK-step predictionrare word recognitionWhispersynthetic audio adaptationreward revocationgreedy decoding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Trie-based biasing rewards partial hypotheses that could lead to a rare word (like "Bon" for "Bonham"), but when the full word never appears, the reward must be revoked in beam search — an expensive step that fails in greedy decoding. This paper claims that a K-step prediction branch, trained on just 10 hours of synthetic audio, lets the model look ahead and check whether the next subword token actually continues the rare word, so the reward can be withheld up front instead of revoked later. The branch attaches to Whisper as an extra decoder layer whose gradients are isolated from the original decoder. On the Singapore English NSC Part 2 set, the method reduces word error rate from 30.86% to 12.19% and outperforms naive trie-based biasing when 100 distractor words are in the bias list. If correct, the approach makes contextual biasing cheaper and compatible with greedy decoding.

Core claim

The paper's central claim is that the indicator function that decides which partial hypotheses get a biasing bonus can be upgraded from "does this prefix appear in a biased word?" to "does this prefix appear in a biased word, and do the top-mu tokens of a multi-step look-ahead continue that word?" Concretely, Whisper is fine-tuned with an extra decoder layer placed before the last layer that outputs the next-next token ("c d e" given "a b c"), and gradients from this layer are prevented from back-propagating into the original decoder. At decoding time, a bonus is applied to a prefix only if both conditions hold, so a prefix like "[Bon]" is not rewarded when the model predicts the next token

What carries the argument

The K-step prediction branch: an extra decoder layer added just before Whisper's last decoder layer, trained on synthetic audio to forecast the token after the current one (and generally K-1 future tokens) in a single pass. Its output is used by the modified indicator 1'(y1:n), which grants a trie-bias bonus only when the prefix matches a biased word and the top-mu predicted future tokens continue that word. This gate replaces the reward-revocation step of standard trie-based biasing and works without beam search.

Load-bearing premise

The whole gain rests on the bet that a look-ahead branch trained on 10 hours of synthetic audio can accurately predict the next subword token for real, differently-accented speech containing rare words it never heard.

What would settle it

Run the K-step branch on the real NSC test set and compute how often the correct continuation of a rare-word prefix falls outside the top-mu=10 predicted tokens; if that miss rate is high, the gate will withhold bonuses from correct prefixes or grant them to wrong ones, and the WER drop should shrink or reverse.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Trie-based biasing can run with greedy decoding; the reward-revocation step of beam search is no longer required.
  • With 100 distractors in the bias list, the method reaches 12.64% WER versus 14.57% for naive trie biasing, showing more reliable rare-word selection under ambiguity.
  • Only 10 hours of synthetic audio (with a different accent from the test data) is enough to make the look-ahead branch useful on real test speech.
  • The approach transfers from whisper-small to whisper-large-v2, improving contextual biasing in both cases (Fig. 3).
  • K-step prediction adds less computational overhead than increasing beam size; the cost increase relative to beam-1 decoding without KP stays modest (Fig. 4).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: If the look-ahead predictor proves reliable on other accents and domains, the same gating idea could replace beam-search-based revocation in streaming ASR, where greedy decoding is the norm.
  • Inference: The paper leaves the future-token predictor's accuracy unmeasured; a direct measure of top-mu recall on real test audio would show how much of the gain is due to the gate versus the synthetic-data fine-tuning.
  • Inference: The benefit of the gate should grow for rare words with long, confusable prefixes, since the look-ahead resolves ambiguity earlier; this suggests a testable extension varying word length and distractor similarity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a K-step prediction extension to Whisper for trie-based contextual biasing. An extra decoder layer, fine-tuned on 10h of synthetic NSC-Part-2 speech, is used to predict one future subword token; during decoding, trie biasing rewards are given only if the top-μ predicted future tokens can continue a biased rare word, thereby avoiding the reward-revocation step of standard trie biasing. On the NSC-Part-2 test set with whisper-small, the method reports WER reductions from 30.86% to 12.19% and from 14.57% to 12.64% at N=100 distractors compared with naive trie biasing. The authors also report a whisper-large-v2 result in Fig. 3 and an ablation on beam size and computational cost in Fig. 4.

Significance. If validated, the proposal is a practically useful contribution: it is an inference-time biasing method that is compatible with greedy decoding, avoids beam-search-only reward revocation, and requires only a small amount of synthetic training data. The paper has concrete strengths: it uses the public NSC-Part-2 corpus with released synthetic audio, reproduces several baselines in a common setup, and includes a computational-cost analysis. However, the central mechanism — the K-step prediction branch — is never evaluated on its own, and the main comparison omits the reward-revocation method the paper claims to replace. The evidence as presented is therefore suggestive but not yet conclusive.

major comments (4)
  1. [§2.4, Fig. 2, Eq. (10)] The gate 1'(y1:n) replaces the trie indicator using 'tokens with the top μ highest probabilities in the K-step prediction.' This gate is the entire mechanism that removes reward revocation, yet the paper reports no standalone measure of the K-step branch's accuracy, precision/recall, or calibration. Critically, the branch is trained on (audio, correct-prefix) pairs but at inference is invoked on possibly wrong partial hypotheses, which are out-of-distribution for the branch. Without an oracle-gate ablation or an analysis of how often the true continuation is in the top-μ set when the prefix is wrong, the improvement in Table 1 cannot be attributed to the K-step gate rather than to the fine-tuned base model or to a coarse bias effect.
  2. [Table 1, §4.1] The paper's core claim is that K-step prediction 'avoids the revocation step entirely' and is an 'alternative to reward revocation.' However, the trie-based biasing baseline is not described as including reward revocation, and no experiment compares against beam-search decoding with reward revocation. §4.1 studies beam size but does not implement the revocation baseline. Without this comparison, the paper does not show that the proposed method matches or exceeds the method it replaces, either in WER or in computational cost. This is a load-bearing omission for the central claim.
  3. [Fig. 2, §2.4, §3.2] The architecture of the K-step branch is underspecified. The text says an extra decoder layer is added 'right before the last decoder layer' to produce the next-next token prediction, and that 'the gradients from the extra decoder layer will not back-propagate to the original decoder,' but also that 'only the encoder weights are frozen.' It is not clear which hidden state feeds the extra layer, how the branch is trained (what loss, what target sequence, whether the branch has its own output projection), how 'K-step' is realized for K=2, or how gradient blocking is implemented. These details are necessary for reproduction and for assessing whether the branch can plausibly generalize to real-audio conditions.
  4. [Table 1, §3.2] All results are single-run numbers without confidence intervals or significance tests. The N=10 comparison between trie biasing and the proposed method is 12.56 vs 12.19 WER, a 0.37-point difference that could be within run-to-run or test-set variability. In addition, the key hyperparameters λ, μ, and K are tuned on the development set, and no sensitivity analysis is reported. At minimum, the authors should provide bootstrap confidence intervals or multiple-seed runs, and a brief sensitivity table for λ, μ, and K, so the reader can judge the stability of the reported gains.
minor comments (6)
  1. [§1] 'Bonhan' appears where 'Bonham' is meant; please fix the typo.
  2. [§2.1, Eq. (1)] P(y1|, x) has an empty conditioning argument; it should be P(y1|x) or P(y1|y0, x) with y0 defined.
  3. [§2.4] With K=2, only one future token is predicted, so the phrase 'tokens with the top μ highest probabilities in the K-step prediction' is ambiguous. Specify whether μ is per future position or over the single next token.
  4. [Table 1] The repeated rows for unadapted and FT baselines across N values make the table harder to read; consider indicating 'independent of N' explicitly. The row for Kwok et al. [24] is also unclear (16.5 with no WER/BWER/UWER breakdown).
  5. [Fig. 3] Consider adding numerical values and, where possible, error bars to the figure; currently the reader cannot read off precise WERs.
  6. [§3.2] No code or detailed hyperparameter table is provided. Releasing the fine-tuning recipe and the K-step branch training details would substantially improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity found; central K-step gate is an independently trained estimator evaluated on held-out data.

full rationale

After walking the derivation chain, I find no step in which a reported prediction is equivalent by construction to an input fit or to a self-citation. The contextual biasing derivation (Eqs. 2-10) is standard algebraic manipulation with scalar hyperparameters α, β (combined into λ); the K-step modification replaces the hard trie-membership indicator 1(y1:n) with a conjunction that additionally requires the top-μ tokens from a separately trained future-token branch to continue the trie path. That branch is trained on 10h of synthetic audio to predict actual future subwords and is then evaluated on a held-out real test set, so the "look-ahead" is a genuine learned estimate rather than a restatement of the indicator. The main reported numbers are held-out WERs after dev-set hyperparameter tuning; λ, μ, K are tuned, not predicted. Self-citations such as [24] supply the synthetic data and a baseline, but are externally reproducible and not used to derive or justify the K-step mechanism. The lack of a reward-revocation baseline and the absence of a standalone K-step accuracy evaluation weaken the empirical support but do not make the argument circular.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central method relies on a new decoder branch and on the reliability of its future-token predictions. The free parameters lambda, mu, and K are fitted on the dev set. The probabilistic derivation borrows standard contextual biasing assumptions, so most of the paper's load is carried by the unproven reliability of the K-step gate under domain mismatch.

free parameters (3)
  • lambda (trie reward) = 3
    Positive log-add bonus for tokens on a biased-word path; tuned on the dev set (Sec. 3.2, Eq. 10).
  • mu (top-K future tokens considered) = 10
    Number of highest-probability K-step predictions that count as continuing a biased word; set in Sec. 3.2.
  • K (look-ahead steps) = 2
    Chosen because Whisper has difficulty predicting tokens for K > 2 steps with the limited synthetic data (Sec. 3.2).
axioms (4)
  • domain assumption Q(y|C,x) = P(y|C,x) and Q(y|not C,x) = P(y|not C,x), with only Q(C|x) increased
    Stated in Sec. 2.2, Eq. 3-4; the entire contextual biasing adjustment assumes the test distribution differs from training only in the prior probability of the bias-list event.
  • domain assumption P(y_n|C,y1:n-1,x) is approximated by P(y_n|y1:n-1,x) times an indicator of the partial hypothesis
    Trie-based biasing's indicator estimator, Eq. 6-7; this is an approximation, not a derivation.
  • domain assumption The K-step branch's top-mu future-token predictions are a reliable proxy for whether the partial hypothesis will complete a biased word
    Core premise of Sec. 2.4 and Fig. 2; if the branch is wrong, the gate over- or under-biases and there is no revocation.
  • ad hoc to paper Gradient blocking from the added decoder layer preserves the original decoder's behavior
    Stated in Fig. 2; no ablation shows whether the original decoder is truly unaffected.
invented entities (1)
  • Extra decoder layer for K-step prediction no independent evidence
    purpose: Predicts future token(s) from the penultimate decoder layer to gate trie biasing rewards
    The paper evaluates the full system on NSC Part 2, but no standalone benchmark of the future-token predictor is provided and no model is released.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Trie-based Biasing using K-step Prediction for Rare Word Recognition." pith.science (2026). https://pith.science/paper/6VBI2BZ4

@misc{pith2026250909196,
  author       = {Pith},
  title        = {Pith review of: Efficient Trie-based Biasing using K-step Prediction for Rare Word Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6VBI2BZ4}},
  note         = {Machine review of arXiv:2509.09196}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Contextual biasing improves rare word recognition of ASR models by prioritizing the output of rare words during decoding. A common approach is Trie-based biasing, which gives "bonus scores" to partial hypothesis (e.g. "Bon") that may lead to the generation of the rare word (e.g. "Bonham"). If the full word ("Bonham") isn't ultimately recognized, the system revokes those earlier bonuses. This revocation is limited to beam search and is computationally expensive, particularly for models with large decoders. To overcome these limitations, we propose adapting ASR models to look ahead and predict multiple steps at once. This avoids the revocation step entirely by better estimating whether a partial hypothesis will lead to the generation of the full rare word. By fine-tuning Whisper with only 10 hours of synthetic data, our method reduces the word error rate on the NSC Part 2 test set from 30.86% to 12.19%.

Figures

Figures reproduced from arXiv: 2509.09196 by Chin Yuen Kwok, Jia Qi Yip.

Figure 1
Figure 1. Figure 1: Overview of incorporating K-step prediction for Trie￾based biasing using Whisper. Given that the input audio says “Bulan” and A) the biased words are “Bulan” and “Bonham”, B) naive Trie-based biasing will bias both tokens “[Bon]” and “[Bu]” with equal rewards. If Whisper wrongly output a higher probability for “[Bon]”, the output token will be “[Bon]”. C) After adapting Whisper to perform K-step prediction… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our Whisper architecture with K-step prediction. Assume K = 2 and only one extra future token (next-next token) is predicted, and assume the decoder has four layers. Given a sequence “a b c d e”, A) the decoder is trained to take in “a b c” to B) generate “b c d” in an auto-regressive manner for next-token prediction. To extend Whisper with two￾step prediction capabilities, C) an extra decoder … view at source ↗
Figure 3
Figure 3. Figure 3: Contextual biasing for whisper-large-v2. Only syn￾thetic data is used for model adaptation. N is the number of distractors added to the biasing words list. KP means K-step prediction [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The effect of beam size on the WER performance and computational cost of Trie-based biasing. KP means K-step prediction. ∆C means the increase in the computational cost relative to decoding with beam size of 1 without KP. N is set to 100. for beam sizes of 1 and 2, but saturates at beam 3. We hypoth￾esize that this is because the large rewards used for Trie-based biasing are tuned for smaller beam sizes an… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 11 linked inside Pith

  1. [1]

    Bulan” and A) the biased words are “Bulan

    Introduction Accurate recognition of rare words is essential for ASR sys- tems, as real-world speech often includes new, domain-specific, or underrepresented terms. However, the rare occurrence of these words in training data limits ASR performance. To ad- dress this, synthetic training data generated via text-to-speech (TTS) systems can be used as a subs...

    Pith/arXiv arXiv 2025

  2. [2]

    Bulan” and the biased words are “Bulan

    Method 2.1. Beam Search Given acoustic observationsxfrom an audio sample and its corresponding transcript represented as a sequence of subword unitsy 1:L = (y1, . . . , yL), end-to-end ASR models are trained to generate the subword units in an autoregressive manner. Specifically, these models estimate the posterior probability of each subword unit conditi...

  3. [3]

    Experiment Setup 3.1. Dataset and metrics Experiments were conducted on the National Speech Corpus Part 2 (NSC-Part-2) [23], a Singapore English dataset compris- ing 13K unique utterances of people asking for directions, in- cluding road names and addresses. A development set of 500 utterances was held out. This dataset was selected because synthetic audi...

  4. [4]

    Initially, the model performs poorly on NSC-Part-2, with a WER of 30.85%

    Results and Discussions Table 1 presents the contextual biasing results for whisper- small. Initially, the model performs poorly on NSC-Part-2, with a WER of 30.85%. Fine-tuning (FT) on real NSC-Part-2 train- ing data substantially improves WER to 10.04%. As real audio training data may be unavailable, synthetic data can be used 1https://github.com/coqui-...

  5. [5]

    Conclusion Our proposed method enhances contextual biasing by enabling ASR models to predict multiple future tokens, eliminating the need for reward revocation in Trie-based biasing. Fine-tuning Whisper with just 10 hours of synthetic data significantly re- duced the WER from 30.86% to 12.19%, demonstrating the ef- fectiveness of our method in improving r...

  6. [6]

    Acknowledgements The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore (https://www.nscc.sg)

  7. [7]

    Keyword-guided adaptation of automatic speech recognition,

    A. Shamsian, A. Navon, N. Glazer, G. Hetz, and J. Keshet, “Keyword-guided adaptation of automatic speech recognition,” arXiv preprint arXiv:2406.02649, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Using synthetic audio to improve the recognition of out-of-vocabulary words in end-to-end asr systems,

    X. Zheng, Y . Liu, D. Gunceler, and D. Willett, “Using synthetic audio to improve the recognition of out-of-vocabulary words in end-to-end asr systems,”ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5674–5678, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:227126398

    2021

  9. [9]

    Imple- menting contextual biasing in gpu decoder for online asr,

    I. Nigmatulina, S. Madikeri, E. Villatoro-Tello, P. Motli ˇcek, J. Zuluaga-Gomez, K. Pandia, and A. Ganapathiraju, “Imple- menting contextual biasing in gpu decoder for online asr,”arXiv preprint arXiv:2306.15685, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Adaptive contextual biasing for transducer based streaming speech recognition,

    T. Xu, Z. Yang, K. Huang, P. Guo, A. Zhang, B. Li, C. Chen, C. Li, and L. Xie, “Adaptive contextual biasing for transducer based streaming speech recognition,”arXiv preprint arXiv:2306.00804, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Selective Biasing with Trie- based Contextual Adapters for Personalised Speech Recognition using Neural Transducers,

    P. Harding, S. Tong, and S. Wiesler, “Selective Biasing with Trie- based Contextual Adapters for Personalised Speech Recognition using Neural Transducers,” inProc. INTERSPEECH 2023, 2023, pp. 256–260

    2023

  12. [12]

    Contextual adapters for personalized speech recognition in neural transducers,

    K. M. Sathyendra, T. Muniyappa, F.-J. Chang, J. Liu, J. Su, G. P. Strimel, A. Mouchtaris, and S. Kunzmann, “Contextual adapters for personalized speech recognition in neural transducers,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8537– 8541

    2022

  13. [13]

    Effective Training of Attention-based Contextual Biasing Adapters with Synthetic Audio for Personalised ASR,

    B. Naowarat, P. Harding, P. D’Alterio, S. Tong, and B. Awwad Shiekh Hasan, “Effective Training of Attention-based Contextual Biasing Adapters with Synthetic Audio for Personalised ASR,” in Proc. INTERSPEECH 2023, 2023, pp. 1264–1268

    2023

  14. [14]

    Can contextual biasing remain effective with whisper and gpt-2?

    G. Sun, X. Zheng, C. Zhang, and P. C. Woodland, “Can contextual biasing remain effective with whisper and gpt-2?”

  15. [15]

    Contextualizing asr lattice rescoring with hybrid pointer network language model,

    D.-R. Liu, C. Liu, F. Zhang, G. Synnaeve, Y . Saraf, and G. Zweig, “Contextualizing asr lattice rescoring with hybrid pointer network language model,”arXiv preprint arXiv:2005.07394, 2020

    Pith/arXiv arXiv 2005

  16. [16]

    Can Contex- tual Biasing Remain Effective with Whisper and GPT-2?

    G. Sun, X. Zheng, C. Zhang, and P. C. Woodland, “Can Contex- tual Biasing Remain Effective with Whisper and GPT-2?” inProc. INTERSPEECH 2023, 2023, pp. 1289–1293

    2023

  17. [17]

    Phoneme-aware encoding for prefix-tree-based contextual asr,

    H. Futami, E. Tsunoo, Y . Kashiwagi, H. Ogawa, S. Arora, and S. Watanabe, “Phoneme-aware encoding for prefix-tree-based contextual asr,” inICASSP 2024-2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 641–10 645

    2024

  18. [18]

    G2g: Tts-driven pronunciation learning for graphemic hybrid asr,

    D. Le, T. Koehler, C. Fuegen, and M. L. Seltzer, “G2g: Tts-driven pronunciation learning for graphemic hybrid asr,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6869–6873

    2020

  19. [19]

    Deep shallow fusion for rnn-t personalization,

    D. Le, G. Keren, J. Chan, J. Mahadeokar, C. Fuegen, and M. L. Seltzer, “Deep shallow fusion for rnn-t personalization,” in2021 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021, pp. 251–257

    2021

  20. [20]

    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. ISCA, pp. 257–261, 2024

    2024

  21. [21]

    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, 2022. [Online]. Available: https://api.semanticscholar. org/CorpusID:252923993

    2022

  22. [22]

    Improved alignment for score combination of rnn-t and ctc decoder for online decod- ing,

    C. Y . Kwok, J. Q. Yip, and E. S. Chng, “Improved alignment for score combination of rnn-t and ctc decoder for online decod- ing,” inInternational Conference on Text, Speech, and Dialogue. Springer, 2024, pp. 70–80

    2024

  23. [23]

    Continual learning op- timizations for auto-regressive decoder of multilingual asr sys- tems,

    C. Y . Kwok, J. Q. Yip, and E. S. Chng, “Continual learning op- timizations for auto-regressive decoder of multilingual asr sys- tems,”arXiv preprint arXiv:2407.03645, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Continual learning with embedding layer surgery and task- wise beam search using whisper,

    ——, “Continual learning with embedding layer surgery and task- wise beam search using whisper,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 140–146

    2024

  25. [25]

    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,” inInterspeech, 2019. [Online]. Available: https: //api.semanticscholar.org/CorpusID:202716857

    2019

  26. [26]

    Deep context: End-to-end contextual speech recognition,

    G. Pundak, T. N. Sainath, R. Prabhavalkar, A. Kannan, and D. Zhao, “Deep context: End-to-end contextual speech recognition,”2018 IEEE Spoken Language Technology Workshop (SLT), pp. 418–425, 2018. [Online]. Available: https://api. semanticscholar.org/CorpusID:51942169

    2018

  27. [27]

    Look-ahead techniques for fast beam search,

    S. Ortmanns and H. Ney, “Look-ahead techniques for fast beam search,”Computer Speech & Language, vol. 14, no. 1, pp. 15–32, 2000

    2000

  28. [28]

    A new verification- based fast match approach to large vocabulary speech recogni- tion,

    F. Liu, M. Afify, H. Jiang, and O. Siohan, “A new verification- based fast match approach to large vocabulary speech recogni- tion,” inProc. Eurospeech 2001, 2001, pp. 851–854

    2001

  29. [29]

    We adapt whisper-small and whisper-large-v2 with vanilla fine-tuning (FT) for 2 epochs and set the train batch size to 6

    toolkit. We adapt whisper-small and whisper-large-v2 with vanilla fine-tuning (FT) for 2 epochs and set the train batch size to 6. We set the learning rate to0.005. We use the AdamW op- timizer [30] with a variant3 of the ReduceLROnPlateau learning rate (LR) scheduler. The encoder weights are frozen to prevent overfitting [31, 32]. We sweep through the hy...

  30. [30]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,”arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  31. [31]

    Building the singapore english national speech corpus,

    J. X. Koh, A. Mislan, K. Khoo, B. Ang, W. Ang, C. Ng, and Y . Tan, “Building the singapore english national speech corpus,” Malay, vol. 20, no. 25.0, pp. 19–3, 2019

    2019

  32. [32]

    Asr model adaptation for rare words using synthetic data generated by multiple text- to-speech systems,

    C. Y . Kwok, H. Y . Li, and E. S. Chng, “Asr model adaptation for rare words using synthetic data generated by multiple text- to-speech systems,” in2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2023, pp. 1771–1778

    2023

  33. [33]

    Libritts: A corpus derived from librispeech for text- to-speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text- to-speech,”arXiv preprint arXiv:1904.02882, 2019

    Pith/arXiv arXiv 1904

  34. [34]

    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,” inInter- national Conference on Machine Learning. PMLR, 2021, pp. 5530–5540

    2021

  35. [35]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92),” 2019. [Online]. Available: https: //api.semanticscholar.org/CorpusID:213060286

    2019

  36. [36]

    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

    Pith/arXiv arXiv 2021

  37. [37]

    Speechbrain: A general-purpose speech toolkit,

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y . Gao, R. D. Mori, and Y . Bengio, “Speechbrain: A general-purpose speech toolkit,” 2021. [Online]. Available: https://arxiv.org/abs/2106.04624

    Pith/arXiv arXiv 2021

  38. [39]

    Low resource language adaptation using two-stage regularization for multilingual asr,

    C. Y . Kwok, J. Q. Yip, and E. S. Chng, “Low resource language adaptation using two-stage regularization for multilingual asr,” in2024 International Conference on Asian Language Processing (IALP). IEEE, 2024, pp. 332–337

    2024

  39. [40]

    Extending whisper for emotion prediction using word-level pseudo labels,

    C. Y . Kwok, S. Li, J. Q. Yip, C. Chu, T. Kawahara, and E. S. Chng, “Extending whisper for emotion prediction using word-level pseudo labels,” inICASSP 2025-2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  40. [2023]

    Available: https://arxiv.org/abs/2306.01942

    [Online]. Available: https://arxiv.org/abs/2306.01942

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.