Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Phonetically-Augmented Discriminative Rescoring for Voice Search Error Correction

T0 review · 3 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 Adding a phonetic search step and a discriminative rescorer to a strong ASR system reduces word error rate on movie-title voice queries by 4.4% to 7.6% relative.

desk verdict A solid industrial integration of phonetic search and discriminative rescoring with real WER gains, but the evaluation dodges the paper's own motivating case of rare titles. read the letter →

arxiv 2506.06117 v1 pith:EH57CDV6 submitted 2025-06-06 cs.CL cs.AIcs.IR

Christophe Van Gysel , Maggie Wu , Lyan Verwimp , Caglar Tirkaz , Marco Bertola , Zhihong Lei , Youssef Oualil This is my paper
classification cs.CLcs.AIcs.IR
keywords ASRerrorcorrectionphoneticsearchdiscriminativerescoringvoicemovietitlerecognitionhiddenMarkovmodelweightedfinite-statetransducersminimumwordrate
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 aims to show that a phonetic correction layer can fix voice-search errors for movie titles that a first-pass end-to-end ASR model never considers. It proposes a two-stage system: a phone-to-text search that generates phonetic alternatives over an HMM built from the ASR's own lexical and language resources, followed by a discriminative rescorer that fuses those alternatives with the ASR N-best list. On two movie-voice-search benchmarks the combined system improves relative word error rate by 4.4% to 7.6% over a baseline that already includes LM rescoring and acoustic-model fusion, and it does not hurt general-query performance. The significance is that correction candidates come from phonetics rather than from a token-to-token model trained on ASR errors, so the approach needs only a small amount of paired audio-text data and reuses existing ASR components.

What carries the argument

The central mechanism is a hidden Markov model that treats the ASR system's phonetic transcription as an observed string and a word-based phonetic language model as the state space, allowing insertions, deletions, and substitutions through learned confusion probabilities. The hidden states come from composing a unigram lexicon finite-state transducer with a word-level finite-state acceptor; the emission probabilities and insertion probability are estimated by aligning phonetic hypotheses to ground-truth transcriptions. A dynamic-programming search over this HMM produces the top-M phonetic alternatives. Those alternatives are then merged with the top-N ASR hypotheses and scored by a discriminative rescorer trained with a minimum-word-error-rate loss on about 80,000 utterances, using acoustic, phonetic, language-model, and source features.

What would settle it

Take a held-out set of movie titles that are deliberately absent from the phonetic LM graph used by PTT, with their pronunciations also removed. If the system's WER on that set does not improve after adding PTT and the rescorer, the claimed gains depend on vocabulary coverage rather than on the correction machinery itself; conversely, if gains persist when titles are present only in the phonetic graph and never in the ASR's training data, the method's core premise is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that the correct transcription for many misrecognized movie titles is recoverable from the audio's phonetic sequence once the search space is driven by a phonetic language model instead of the E2E model's word posterior. The evidence is the oracle WER of the combined ASR-plus-PTT candidate list, which on the three test sets is substantially lower than the final WER, meaning the rescorer's main job is selection rather than generation. The paper reports a 4.40% relative WER improvement on 'play $TITLE' queries and a 7.55% improvement on title-only queries over the strongest baseline, attributed to the ability to propose hypotheses outside the first-pass N-best list.

Load-bearing premise

The phonetic language model graph must contain the correct movie title with its pronunciation for every test query; otherwise the phonetic search cannot propose it.

Editorial extensions

If this is right

  • The same two-stage design applies to other rare-entity voice search domains, such as music titles, app names, or contact names, without retraining the ASR model.
  • New entities can be handled by updating the phonetic LM graph (lexicon and language model), rather than by collecting new paired audio-text data.
  • The rescorer's small training requirement (~80k utterances) makes the approach practical for on-device or frequently updated deployments.
  • Ablations show LM features are essential; without them the rescorer cannot pick among phonetically close candidates, so future work should concentrate on richer LM scoring.

Reading between the lines

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

  • If the gains hold across a wider variety of queries, phonetic correction could become a general post-processing layer for any entity-centric ASR, not just media search.
  • The method's dependence on vocabulary coverage implies that a hybrid with a token-to-token model might handle true out-of-vocabulary titles, since PTT alone cannot propose unseen strings.
  • Prepending 'play' changes which feature group matters (phonetic vs acoustic), suggesting the optimal correction strategy depends on the query's left context; a context-aware rescorer might improve both regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes a two-stage phonetic correction system for end-to-end ASR in voice search. After a Conformer CTC + TLG + external LM rescoring + acoustic model fusion pipeline produces an N-best list, a phone-to-text (PTT) component uses a phonetic HMM over a word-based LM graph G to generate alternative word sequences, and a discriminative rescorer combines the PTT alternatives with the ASR top hypothesis using features from phonetics, acoustics, and language models. Experiments on movie-title voice search show relative WER reductions of 4.4% (PlayMovie) and 7.55% (VerblessMovie) over a strong baseline, with no degradation on general VA queries. The paper also reports oracle WERs of the combined K-best lists and an ablation study of feature groups.

Significance. The work is practically relevant because it avoids training a token-to-token correction model, which requires very large audio-text corpora and is tightly coupled to the ASR system. Instead, it reuses the existing external LM, monophone acoustic model, and pronunciation lexicon, and trains only a lightweight rescorer on about 80k utterances. The ablation study supports the necessity of both LM and acoustic/phonetic features, and the oracle analysis helps diagnose the reachable upper bound. If the reported gains are reliable, this is a practical, deployable add-on for voice-search error correction. The system description is clear and the decomposition between candidate generation and rescoring is a useful design pattern.

major comments (3)
  1. [§3.2, §3.1, §1] The evaluation uses top-K most popular movie titles, which are likely well represented in the n-gram LM used as graph G (and possibly in the E2E training data as well). This does not test the paper's motivating scenario of recent or infrequent titles that are under-represented in the E2E system. Since PTT can only propose titles present in G, the reported improvements are conditional on the correct title being in the LM. The paper should report the proportion of test utterances whose correct title is covered by G, and ideally add an evaluation on rare or post-cutoff titles, or at least stratify results by title frequency/popularity. As written, the headline claim does not establish the emerging-entity case.
  2. [Table 4] The paper reports relative WER improvements of 4.4% and 7.55% (absolute 0.21 and 0.43 percentage points) on test sets of 2k and 4k utterances, but provides no confidence intervals or statistical significance tests. Given the small absolute differences, the improvements could in principle arise from noise. The authors should report bootstrap confidence intervals or a pairwise significance test (e.g., matched-pair sentence error or MAPSSWE) to support the claim.
  3. [§4 (oracle WERs)] The oracle WERs of the phonetically-augmented K-best lists (4.0, 2.94, 3.47) are substantially lower than the final system WERs (5.92, 4.58, 5.37). On GeneralVA, the oracle is 4.0 while the final system remains at 5.92, indicating that the rescorer does not exploit available oracle hypotheses. The paper should discuss this gap, especially on GeneralVA, and analyze whether it reflects a limitation of the rescorer features, the training data, or the rescoring objective. This is important for understanding the approach's limits.
minor comments (5)
  1. [§2.2, Eq. (1)] The notation f(h_i) = x_i · w^⊺ is ambiguous; use f(h_i) = x_i^⊺ w (or define w as a column vector and write f(h_i) = x_i · w as a dot product).
  2. [§3.1] It is unclear why the corrections system uses only N=1 from the ASR N-best list when the preceding amfusion system uses N=25, while §2.1.3 states that PTT selects a phonetic transcription from the ASR N-best. Clarify the relationship between the N used for PTT input and the N used in the final rescoring, and motivate this choice.
  3. [Table 1b and §2.1.2] The 'Phonetic distance C_Phon(h | h*)' is not explicitly defined in terms of the confusion probabilities P(o|o'). Specify how the distance is computed (e.g., negative log-likelihood of the alignment, phone error rate, etc.).
  4. [Figure 2] The ablation study is presented only as a figure; the text gives qualitative statements (e.g., 'most of the gains vanish when LM features are removed') but no numerical WER values. Add the actual relative WER changes in the text or figure caption for reproducibility.
  5. [Throughout] There are minor typographical issues, for example 'minimium WER objective' in §2.2 should be 'minimum WER objective' and 'the the' appears in one place in §2.1.3. A careful copyedit is recommended.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: held-out WER evaluation and a non-trivial oracle gap show the claimed gains are not forced by construction.

full rationale

The paper's central claim is an empirical WER improvement on held-out benchmarks (Table 4), produced by a two-stage system whose parameters (phonetic confusion probabilities, rescorer weights) are fit on development data separate from the test utterances. No equation in the paper defines the reported improvement in terms of a fitted quantity: the confusion model estimates P(o|o') from NW alignments of dev data (§2.1.2), and the rescorer minimizes a min-WER loss on a separate 80k-utterance training set (§2.2, §3.3). The final result is measured on GeneralVA, PlayMovie, and VerblessMovie test sets (§3.2). The reported oracle WERs of the combined K-best lists (4.0, 2.94, 3.47) versus final system WERs (5.92, 4.58, 5.37) demonstrate that the rescorer is genuinely selecting among non-equivalent hypotheses, not returning its training targets. The main in-scope concern is that PTT uses the same n-gram LM as graph G (§3.1: 'PTT uses the n-gram LM from (1) as graph G'), and the rescorer also consumes LM features; this creates shared knowledge between candidate generation and reranking, but it does not make the predicted WER equal to the LM score by construction. If a correct movie title is absent from G, PTT cannot propose it; the paper does not measure coverage of G over the test titles. That is a limitation of external validity, not circularity. Self-citations (e.g., Lei et al. [17] for the AM-fusion baseline, Van Gysel et al. [32] for the entity-centric LM) are cited for components that are independently described and measured; they do not invoke an author-imported uniqueness theorem or smuggle an ansatz as an external fact. No step in the derivation chain reduces to its own input, so the appropriate finding is no significant circularity; score 1 reflects only the normal presence of in-house baselines and shared LM components, which are not load-bearing circular inputs.

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

The system's performance rests on standard estimation of error-channel probabilities and on the availability of a phonetic LM that covers the target entities. No new physical entities are introduced.

free parameters (4)
  • Phonetic emission probabilities P(o|o') = counts from Needleman-Wunsch alignments on development set
    Used in the HMM emission model (Section 2.1.2); estimated from aligned ASR phonetic hypotheses and ground-truth transcriptions.
  • Insertion probability p_ins = proportion of epsilon alignments in development set
    Controls the likelihood of the insertion sub-graph in the HMM (Section 2.1.2).
  • Feature transformation threshold t = log(10^-7)
    Hand-chosen threshold for the binary feature in Table 2 row 5; no justification is given.
  • Rescorer weight vector w = learned, not reported
    Optimized on 80,000 utterances via the minimum WER objective (Section 2.2). These are trained parameters, but they are fitted to data.
assumptions (5)
  • domain assumption The phonetic LM graph G (word vocabulary and pronunciation entries) contains the correct movie title for every test query.
    PTT can only propose words in L and G (Sections 2.1.1 and 3.1). If a title is out of vocabulary or its pronunciation is missing, correction fails on the motivating rare-entity case.
  • domain assumption The Monophone AM's phonetic transcription is close enough to the ground-truth phones that the correct title is reachable by a bounded number of edit operations in the HMM.
    The HMM edit model in Section 2.1 assumes the observed ASR phonetic sequence differs from the reference via insertions, deletions, and substitutions with probabilities estimated from development data. Large phone-level errors break this assumption.
  • domain assumption Confusion probabilities estimated on the development set transfer to the test distribution without adaptation.
    Emission and insertion probabilities are estimated from a set of development utterances (Section 2.1.2) and applied to test queries without adaptation.
  • domain assumption The minimum-WER loss and hand-built features in the rescorer generalize from 80,000 training utterances to held-out queries.
    The rescorer is trained on a mixture of human-graded and synthetic queries (Section 3.3) and evaluated on separate test sets; this assumes the feature representation is sufficient for selection.
  • standard math Standard algorithms (Viterbi, Needleman-Wunsch, WFST composition) are used without proof.
    The paper relies on established, unproved background results in Sections 2.1.1 and 2.1.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phonetically-Augmented Discriminative Rescoring for Voice Search Error Correction." pith.science (2026). https://pith.science/paper/EH57CDV6

@misc{pith2026250606117,
  author       = {Pith},
  title        = {Pith review of: Phonetically-Augmented Discriminative Rescoring for Voice Search Error Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EH57CDV6}},
  note         = {Machine review of arXiv:2506.06117}
}
read the original abstract

End-to-end (E2E) Automatic Speech Recognition (ASR) models are trained using paired audio-text samples that are expensive to obtain, since high-quality ground-truth data requires human annotators. Voice search applications, such as digital media players, leverage ASR to allow users to search by voice as opposed to an on-screen keyboard. However, recent or infrequent movie titles may not be sufficiently represented in the E2E ASR system's training data, and hence, may suffer poor recognition. In this paper, we propose a phonetic correction system that consists of (a) a phonetic search based on the ASR model's output that generates phonetic alternatives that may not be considered by the E2E system, and (b) a rescorer component that combines the ASR model recognition and the phonetic alternatives, and select a final system output. We find that our approach improves word error rate between 4.4 and 7.6% relative on benchmarks of popular movie titles over a series of competitive baselines.

Figures

Figures reproduced from arXiv: 2506.06117 by the authors.

Figure 1
Figure 1. Diagram representing a single step in our HMM. The dotted rectangle represents the component responsible for allowing symbols to occur in the observation, with probabil￾ity pins, that are missing from the Phonetic LM. The dashed, rounded rectangle corresponds to a state in the Phonetic LM, with arcs labeled with phonetic units. The rectangular states are emitting states and are annotated with their emission prob￾abi… view at source ↗
Figure 2
Figure 2. Ablation study demonstrating the rel. WER in￾crease that occurs when removing primary feature groups (see Table 1b) on the PlayMovie (circles) and VerblessMovie (stripes) test sets [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages

  1. [1]

    Phonetically-Augmented Discriminative Rescoring for Voice Search Error Correction

    Introduction Modern Automatic Speech Recognition (ASR) systems consist of an end-to-end (E2E) neural model that outputs transcribed text directly given audio as input. E2E ASR models are trained using paired audio-text samples [1] that are expensive to obtain. Digital media players allow users to playback digital con- tent (e.g., movies) on a television, ...

    work page Pith review arXiv 2024

  2. [2]

    Sherlock Holmes

    Methodology Our approach operates within the decoding framework pro- posed by Lei et al. [17], where a subword-based Conformer CTC ASR model [18, 19] is fused with a word-based external LM [20] during the 1st pass, and a phoneme-based AM (re- ferred to as ”Monophone AM”) is used during a 2nd pass to rescore an N-best list through forced alignment. The Mon...

    work page

  3. [3]

    play $TITLE

    Experimental set-up 3.1. Systems under comparison We compare the following speech recognition systems where each system builds upon the previous one—each adding a com- ponent that aims to improve spoken entity query accuracy (see Table 3 for an overview): (1) Ourbasesystem, a sub-word (8k units) Conformer CTC model [18] with a TLG graph decoder (usingT co...

    work page 2024

  4. [4]

    Thebasesystem provides a baseline single- pass Conformer-driven graph decoding set-up (§3.1)

    Results Table 4 shows the results of our experiments, where each row represents a system that includes the components in the rows that precede it. Thebasesystem provides a baseline single- pass Conformer-driven graph decoding set-up (§3.1). Second pass methods, which rescore the results of the first pass, im- prove upon the result by taking into account a...

    work page

  5. [5]

    Conclusions We introduced a phonetic correction system that generates pho- netic alternatives (PTT) based on the ASR system’s recogni- tion, followed by a discriminative rescorer that selects the best alternative amongst the ASR and PPT N-best lists. PTT allows to generate correction candidates without the need for large- scale audio-text samples, and avo...

    work page

  6. [6]

    Acknowledgments We thank Leo Liu, Linda Arsenault, Man-Hung Siu, Sameer Badaskar, Stephen Pulman, Takaaki Hori, Thiago Fraga da Silva, and the anonymous reviewers for their comments and feedback

    work page

  7. [7]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl¨uter, and S. Watan- abe, “End-to-end speech recognition: A survey,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, pp. 325–351, 2023

    work page 2023

  8. [8]

    Modeling Spoken Information Queries for Virtual Assistants: Open Problems, Challenges and Opportunities,

    C. Van Gysel, “Modeling Spoken Information Queries for Virtual Assistants: Open Problems, Challenges and Opportunities,” inSI- GIR, 2023, p. 3335–3338

    work page 2023

Show all 41 references
  1. [9]

    Server- side rescoring of spoken entity-centric knowledge queries for virtual assistants,

    Y . Zhang, S. Gondala, T. Fraga-Silva, and C. Van Gysel, “Server- side rescoring of spoken entity-centric knowledge queries for virtual assistants,”International Journal of Speech Technology, vol. 27, no. 2, pp. 367–375, 2024

    2024

  2. [10]

    A dis- criminative entity-aware language model for virtual assistants,

    M. Saebi, E. Pusateri, A. Meghawat, and C. Van Gysel, “A dis- criminative entity-aware language model for virtual assistants,” in Interspeech, 2021, pp. 2032–2036

    2021

  3. [11]

    List of movies released between January and De- cember 2024,

    IMDb, “List of movies released between January and De- cember 2024,” https://www.imdb.com/search/title/?title type= feature&release date=2024-01-01,2024-12-31, 2024

    2024

  4. [12]

    A spelling correction model for end-to-end speech recognition,

    J. Guo, T. N. Sainath, and R. J. Weiss, “A spelling correction model for end-to-end speech recognition,” inICASSP. IEEE, 2019, pp. 5651–5655

    2019

  5. [13]

    Correction of auto- matic speech recognition with transformer sequence-to-sequence model,

    O. Hrinchuk, M. Popova, and B. Ginsburg, “Correction of auto- matic speech recognition with transformer sequence-to-sequence model,” inICASSP. IEEE, 2020, pp. 7074–7078

    2020

  6. [14]

    Transformer-based model for asr n-best rescoring and rewriting,

    I. E. Kang, C. Van Gysel, and M.-H. Siu, “Transformer-based model for asr n-best rescoring and rewriting,” inInterspeech, 2024, pp. 3505–3509

    2024

  7. [15]

    Transformer based deliberation for two-pass speech recognition,

    K. Hu, R. Pang, T. N. Sainath, and T. Strohman, “Transformer based deliberation for two-pass speech recognition,” inSLT. IEEE, 2021, pp. 68–74

    2021

  8. [16]

    Improving deliberation by text-only and semi-supervised training,

    K. Hu, T. N. Sainath, Y . He, R. Prabhavalkar, T. Strohman, S. Ma- vandadi, and W. Wang, “Improving deliberation by text-only and semi-supervised training,” inInterspeech, 2022, pp. 4940–4944

    2022

  9. [17]

    Recognizing long-form speech using streaming end-to-end models,

    A. Narayanan, R. Prabhavalkar, C.-C. Chiu, D. Rybach, T. N. Sainath, and T. Strohman, “Recognizing long-form speech using streaming end-to-end models,” inASRU, 2019, pp. 920–927

    2019

  10. [18]

    Application- agnostic language modeling for on-device asr,

    M. Nußbaum-Thom, L. Verwimp, and Y . Oualil, “Application- agnostic language modeling for on-device asr,” inACL Industry Track, 2023, pp. 268–275

    2023

  11. [19]

    Confusion-based query expansion for oov words in spoken document retrieval,

    B. Logan and J. M. V . Thong, “Confusion-based query expansion for oov words in spoken document retrieval,” inICSLP, 2002, pp. 1997–2000

    2002

  12. [20]

    An empirical study of confu- sion modeling in keyword search for low resource languages,

    M. Saraclar, A. Sethy, B. Ramabhadran, L. Mangu, J. Cui, X. Cui, B. Kingsbury, and J. Mamou, “An empirical study of confu- sion modeling in keyword search for low resource languages,” in ASRU. IEEE, 2013, pp. 464–469

    2013

  13. [21]

    Discrimi- native language modeling with conditional random fields and the perceptron algorithm,

    B. Roark, M. Saraclar, M. Collins, and M. Johnson, “Discrimi- native language modeling with conditional random fields and the perceptron algorithm,” inACL, 2004, pp. 47–54

    2004

  14. [22]

    Discriminative rescoring based on minimization of word errors for transcribing broadcast news

    A. Kobayashi, T. Oku, S. Homma, S. Sato, T. Imai, and T. Takagi, “Discriminative rescoring based on minimization of word errors for transcribing broadcast news.” inInterspeech, 2008, pp. 1574– 1577

    2008

  15. [23]

    Acoustic model fusion for end-to-end speech recognition,

    Z. Lei, M. Xu, S. Han, L. Liu, Z. Huang, T. Ng, Y . Zhang, E. Pusateri, M. Hannemann, Y . Deng, and M.-H. Siu, “Acoustic model fusion for end-to-end speech recognition,” inASRU, 2024, pp. 1–7

    2024

  16. [24]

    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,” inICML, 2006, pp. 369–376

    2006

  17. [25]

    Conformer: Convolution-augmented transformer for speech recognition,

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

    2020

  18. [26]

    CTC variations through new WFST topologies,

    A. Laptev, S. Majumdar, and B. Ginsburg, “CTC variations through new WFST topologies,” inInterspeech 2022, 2022, pp. 1041–1045

    2022

  19. [27]

    A generalized hidden markov model with discriminative training for query spelling correction,

    Y . Li, H. Duan, and C. Zhai, “A generalized hidden markov model with discriminative training for query spelling correction,” inSI- GIR, 2012, pp. 611–620

    2012

  20. [28]

    Decoding for channels with insertions, deletions, and substitutions with applications to speech recogni- tion,

    L. Bahl and F. Jelinek, “Decoding for channels with insertions, deletions, and substitutions with applications to speech recogni- tion,”Transactions on Information Theory, vol. 21, no. 4, pp. 404– 411, 1975

    1975

  21. [29]

    A generalized com- position algorithm for weighted finite-state transducers

    C. Allauzen, M. Riley, and J. Schalkwyk, “A generalized com- position algorithm for weighted finite-state transducers.” inInter- speech, 2009, pp. 1203–1206

    2009

  22. [30]

    A comparison of dynamic WFST decoding approaches,

    P. R. Dixon, C. Hori, and H. Kashioka, “A comparison of dynamic WFST decoding approaches,” inICASSP, 2012, pp. 4209–4212

    2012

  23. [31]

    Weighted finite-state trans- ducers in speech recognition,

    M. Mohri, F. Pereira, and M. Riley, “Weighted finite-state trans- ducers in speech recognition,”Comput. Speech Lang., pp. 69–88, 2002

    2002

  24. [32]

    A general method applicable to the search for similarities in the amino acid sequence of two proteins,

    S. B. Needleman and C. D. Wunsch, “A general method applicable to the search for similarities in the amino acid sequence of two proteins,”Journal of molecular biology, vol. 48, no. 3, pp. 443– 453, 1970

    1970

  25. [33]

    Error bounds for convolutional codes and an asymp- totically optimum decoding algorithm,

    A. Viterbi, “Error bounds for convolutional codes and an asymp- totically optimum decoding algorithm,”IEEE Transactions on In- formation Theory, vol. 13, no. 2, pp. 260–269, 1967

    1967

  26. [34]

    Minimum word error rate training for attention-based sequence-to-sequence models,

    R. Prabhavalkar, T. N. Sainath, Y . Wu, P. Nguyen, Z. Chen, C.- C. Chiu, and A. Kannan, “Minimum word error rate training for attention-based sequence-to-sequence models,” inICASSP. IEEE Press, 2018, pp. 4839–4843

    2018

  27. [35]

    Adam: A method for stochastic opti- mization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,” inICLR, 2015

    2015

  28. [36]

    Internal language model estimation for domain-adaptive end-to-end speech recognition,

    Z. Meng, S. Parthasarathy, E. Sun, Y . Gaur, N. Kanda, L. Lu, X. Chen, R. Zhao, J. Li, and Y . Gong, “Internal language model estimation for domain-adaptive end-to-end speech recognition,” inSLT. IEEE, 2021, pp. 243–250

    2021

  29. [37]

    The fixed- size ordinally-forgetting encoding method for neural network lan- guage models,

    S. Zhang, H. Jiang, M. Xu, J. Hou, and L.-R. Dai, “The fixed- size ordinally-forgetting encoding method for neural network lan- guage models,” inACL-IJCNLP, 2015, pp. 495–500

    2015

  30. [38]

    Space-efficient representation of entity-centric query language models,

    C. Van Gysel, M. Hannemann, E. Pusateri, Y . Oualil, and I. Oparin, “Space-efficient representation of entity-centric query language models,” inInterspeech, 2022, pp. 679–683

    2022

  31. [39]

    Error-driven pruning of language models for virtual assis- tants,

    S. Gondala, L. Verwimp, E. Pusateri, M. Tsagkias, and C. Van Gy- sel, “Error-driven pruning of language models for virtual assis- tants,” inICASSP. IEEE, 2021, pp. 7413–7417

    2021

  32. [40]

    Connecting and comparing language model interpolation techniques,

    E. Pusateri, C. Van Gysel, R. Botros, S. Badaskar, M. Hannemann, Y . Oualil, and I. Oparin, “Connecting and comparing language model interpolation techniques,” inInterspeech, 2019, pp. 3500– 3504

    2019

  33. [41]

    Syn- thetic query generation using large language models for virtual as- sistants,

    S. Sannigrahi, T. Fraga-Silva, Y . Oualil, and C. Van Gysel, “Syn- thetic query generation using large language models for virtual as- sistants,” inSIGIR, 2024, pp. 2837–2841

    2024

Pith tools

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