Pith. sign in

REVIEW 3 major objections 5 minor 100 references

The paper claims that feeding a Conformer-Transducer ASR system the encoder-derived context of preceding utterances consistently reduces word and character error rates, by up to 6% relative, on four benchmark tasks.

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 ↗

Adding cross-utterance audio context to Conformer-Transducer ASR reduces WER/CER by 0.5 to 1.1 absolute points on four benchmarks, and a splicing-based batch scheme cuts training time by up to about 19%.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful side-by-side comparison of cross-utterance context methods for Conformer-Transducer, with a real caveat about the undocumented evaluation protocol. the 3 major comments →

arxiv 2508.10456 v1 pith:KJPA2BYX submitted 2025-08-14 eess.AS

Exploring Cross-Utterance Speech Contexts for Conformer-Transducer Speech Recognition Systems

Mingyu Cui , Mengzhe Geng , Jiajun Deng , Chengxi Deng , Jiawen Kang , Shujie Hu , Guinan Li , Tianzi Wang
show 3 more authors
Zhaoqing Li Xie Chen Xunying Liu
This is my paper
classification eess.AS
keywords cross-utterance speech contextConformer-Transducerstreaming ASRnon-streaming ASRutterance splicingelderly speech recognitionGigaSpeechWenetSpeech
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

This paper tries to establish that an end-to-end speech recognizer should hear context beyond the current utterance: the acoustic, prosodic, and lexical state carried by the immediately preceding utterances helps decide what is being said now. It compares four fusion strategies for Conformer-Transducer (C-T) models — concatenating raw input audio, concatenating previous encoder embeddings, projecting pooled previous encoder embeddings, and a new chunked-window variant — and reports that the best contextual system beats its no-context baseline on all four datasets, across English, Mandarin, and Cantonese, with statistically significant average reductions up to 0.9–1.1% absolute (6.0%/5.4% relative) WER/CER in pre-training and 0.51–0.98% absolute (2.0–3.4% relative) after fine-tuning to elderly speech. The paper also argues for a design rule: non-streaming C-T systems should concatenate full previous-utterance encoder embeddings, while streaming C-T systems should use chunk-based contextual embeddings. A batch-training scheme that splices contiguous utterances inside minibatches preserves the left-to-right context order while cutting training time by up to 16.6%.

Core claim

The central discovery is that the form of the context representation, not merely its presence, determines the gain. Letting the current utterance attend to previous utterances through the same Conformer encoder — by concatenating the previous utterance's encoder embeddings (with stop-gradient) into the key/value stream at each layer — yields the largest gains for non-streaming models, outperforming both raw audio concatenation and compressed pooling projection. For streaming models, the same consistency principle favors the chunk-based approach: mask matrices restrict cross-utterance access to fixed-size sliding windows aligned with the streaming chunk structure. On GigaSpeech and WenetSpeec

What carries the argument

The load-bearing mechanism is the contextual key/value augmentation inside each Conformer encoder block: the current utterance's query attends to keys and values built from previous utterances, with the previous representation detached by a stop-gradient and fused as $X^{l,i}_{\mathrm{FFN}} \oplus \mathrm{SG}(H^{l,i-1})$. The three variants differ in how the previous context is compressed before fusion — direct concatenation, attention-pooled low-rank projection, or chunk-masked windows controlled by $M^l_{\mathrm{prev}}$ and $M^l_{\mathrm{cur}}$ — and in whether the context is full-utterance or truncated. The supporting engineering contribution is a minibatch scheme that splices neighbourin

Load-bearing premise

The load-bearing premise is that evaluation decodes each recording's utterances in true chronological order with cross-utterance context reset only at recording boundaries; the paper does not document whether the GigaSpeech and WenetSpeech test harnesses actually guarantee that, and shuffled-order decoding would shrink or erase the reported gains.

What would settle it

Re-run the GigaSpeech DEV/TEST and WenetSpeech TEST evaluations with utterances shuffled within recordings, or decoded independently, while keeping the same models; if the WER/CER advantage of the contextual systems over the baselines disappears or reverses, the reported context gains depend on evaluation order rather than on the fusion mechanism. A simpler audit is to inspect the segment-ordering metadata of the official test sets to confirm the context resets match recording boundaries.

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

If this is right

  • Any Conformer-Transducer system can be upgraded with cross-utterance context by changing only the encoder's attention inputs; the reported gains are 0.9–1.1% absolute WER/CER on 1000-hour English and Mandarin pre-training.
  • The design rule is actionable: choose full previous-utterance encoder embedding concatenation for offline ASR, and chunk-based encoder embeddings with up to 200 frames of previous context for streaming ASR.
  • Contextual gains survive domain transfer: pre-trained contextual C-T models fine-tuned to low-resource elderly English and Cantonese keep statistically significant WER/CER reductions.
  • The utterance-splicing batch scheme recovers most of the training-time cost of context contiguity, speeding training by roughly 8–19% across the compared systems.
  • Inference cost stays modest: the contextual fusion methods add only a few percent to real-time factor on the tested tasks.

Where Pith is reading between the lines

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

  • If the finding generalizes, single-utterance training leaves a cheap signal unused; injecting prior-utterance audio context into self-supervised speech foundation models could improve them without adding text-side biasing or larger data.
  • The streaming/non-streaming split suggests a broader design principle: context fusion should be matched to the temporal granularity of the host attention mask — a chunked recognizer benefits from chunked context, a full-context recognizer from full context.
  • The reported gains presuppose that evaluation utterances are decoded in their original chronological order; if a deployment decodes segments independently or shuffled, the benefit may shrink. This is testable by re-decoding the GigaSpeech and WenetSpeech test sets in shuffled order.
  • Because the gains persist on clinical-style elderly speech, a natural next step is to test whether cross-utterance context helps pathological speech analysis such as dementia detection, where turn-level acoustic consistency carries diagnostic signal; the paper does not test that.
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

3 major / 5 minor

Summary. The paper investigates four methods for incorporating cross-utterance speech context into Conformer-Transducer (C-T) ASR: (A) input audio feature concatenation, (B) cross-utterance Encoder embedding concatenation, (C) cross-utterance Encoder embedding pooling projection, and (D) a chunk-based extension. It also proposes an utterance-splicing batch-training scheme that preserves left-to-right context order while improving GPU utilization. Experiments on GigaSpeech, WenetSpeech, DementiaBank Pitt, and JCCOCC MoCA report statistically significant WER/CER reductions for the best contextual systems, modest RTF increases, and training-time savings. The paper concludes that method (B) is best for non-streaming C-T and method (D) is best for streaming C-T.

Significance. If the reported results hold, this is a useful systematic comparison of cross-utterance context fusion methods for a mainstream E2E architecture, with a practical training-efficiency contribution and a clear design principle for streaming versus non-streaming systems. The paper is careful to report significance tests and efficiency metrics across four datasets and three languages. However, the main empirical claim depends on an undocumented evaluation-time protocol, and no code or models are released; reproducibility is therefore incomplete. The batch-splicing efficiency result is less protocol-dependent and is a solid practical contribution.

major comments (3)
  1. [Section IV-B and Figure 3 caption] The caption says context propagation is reset at recording boundaries during training 'and also evaluation,' but the paper never specifies whether test segments are decoded in chronological order within each recording. For GigaSpeech DEV/TEST and WenetSpeech TEST, the segment ordering of the official evaluation manifests is not described. If the evaluation harness decodes segments independently or in shuffled order, the reported cross-utterance gains cannot be reproduced. This is load-bearing for every headline WER/CER reduction. Please document the exact test-time segment ordering, verify it against the official dataset manifests, or release code that implements the protocol.
  2. [Tables I, II, IV and abstract] The headline numbers are selected from the best among dozens of configurations (e.g., Table I Sys.17, Table II Sys.16/32) with MAPSSWE at α=0.05 applied to each pairwise comparison and no correction for multiple testing. The overall trend is consistent, which reduces the risk, but the 'statistically significant' claim for the best-of-many systems should be qualified or accompanied by a multiple-comparison correction or a pre-specified primary comparison.
  3. [Section I-C.1] Contribution 1 claims 'the first work to investigate cross-utterance speech contexts modeling approaches for C-T ASR models,' but the Introduction (Section I-B) explicitly acknowledges that [68] previously investigated cross-utterance speech context modeling for Conformer-Transducer systems. This is internally inconsistent. Please revise the novelty claim to distinguish the new chunk-based approach and the complete side-by-side comparison, which are the genuinely novel aspects.
minor comments (5)
  1. [Section V-B1, footnote 6] The text says 'with a significant level of α = 0.056'; this should be α = 0.05.
  2. [Section V-B2] Typo: 'learning rate as 1r-5' should read '1e-5'.
  3. [Abstract and Conclusion] The abstract reports the maximum GigaSpeech reduction as 0.9% absolute, but Table II Sys.16 achieves 1.0% absolute over its streaming baseline. Reconcile the wording, e.g., by specifying that 0.9% refers to the non-streaming system.
  4. [Section VI-A1, item ii] The explanation that method (B) outperforms (C) and (D) because of 'modeling consistency' is plausible but not directly tested; consider softening or adding an ablation that isolates context completeness versus compression.
  5. [Tables I, II] For input audio concatenation (A), only the one-previous-utterance configuration is reported. The conclusion that (A) 'does not perform as well' should be restricted to that configuration.

Circularity Check

0 steps flagged

Minor self-referential design choices (method (B) and L=32 from the authors' [68]), but the headline WER/CER claims are measured against independent in-paper baselines and are not forced by construction.

full rationale

The central claim is empirical: four context fusion methods are defined by explicit equations (Eqs. 10-13), trained, and compared against non-contextual Conformer-Transducer baselines trained in the same paper on standard GigaSpeech, WenetSpeech, DementiaBank Pitt, and JCCOCC MoCA evaluation sets. No fitted parameter is renamed as a prediction, and no equation reduces to its own input. The only self-referential elements are design inheritances from the authors' prior work [68]: Section III-B adopts the cross-utterance Encoder embedding concatenation of [68], and footnote 6 sets the pooling projection dimension L=32 'following the best setting in [68] (Table 2, Sys.13 and 16)'. These are not load-bearing circularity because the pooled-projection variant is not the paper's best method, and the headline gains are computed relative to baselines rather than derived from [68]. The external comparisons to Whisper, XLSR-128, and Wav2vec2.0-Conformer use published system numbers and are independent. The main caveat is a test-protocol assumption, not circularity: the Figure 3 caption states cross-utterance context propagation is reset at recording boundaries 'during contextual C-T model training (and also evaluation)', but the paper does not document whether the official GigaSpeech/WenetSpeech evaluation harness decodes utterances in chronological order. This is a reproducibility/validity risk, not a circular derivation. Hence a low score of 2 for minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

No invented physical or ontological entities appear in the paper. The chunk mask matrices and the pooling projection are design choices built from standard attention and linear algebra operations. The main uncharged assumptions are the standard RNN-T/Conformer baseline, the empirical value of adjacent-utterance context, and the unadjusted use of MAPSSWE across many comparisons.

free parameters (4)
  • Number of previous utterances M = 1, 2, or 3; best is 3 for methods B and C in most settings
    Swept over {1,2,3} in Tables I and II; more context generally gives better WER/CER, so the reported headline numbers use the best setting.
  • Chunk context window size for method D = <=60, <=100, <=160, or <=200 frames; best is usually <=160 or <=200
    Hyperparameter sweep over five context-window sizes; the paper reports the best per configuration rather than a parameter-free choice.
  • Pooling projection dimension L = 32
    Set to L=32 following the authors' prior work [68] Table 2, not re-optimized for GigaSpeech, WenetSpeech, DementiaBank, or JCCOCC.
  • Streaming look-ahead chunk size = 20 frames (200 ms)
    Fixed design choice for the streaming baseline; other look-ahead values are not swept in this paper.
axioms (3)
  • standard math Standard RNN-T decomposition with a Conformer Encoder, as in Equations 1 through 9
    Invoked as the baseline architecture without re-derivation in Section II.
  • domain assumption Adjacent utterances within a recording provide useful acoustic and lexical context for transcription
    This is the central hypothesis tested empirically; it is not derived from first principles.
  • domain assumption MAPSSWE at alpha=0.05 is a valid significance test for each reported comparison
    Statistical method from [93] used per comparison without adjustment for the many configurations tested in Tables I and II.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Cross-Utterance Speech Contexts for Conformer-Transducer Speech Recognition Systems." pith.science (2026). https://pith.science/paper/KJPA2BYX

@misc{pith2026250810456,
  author       = {Pith},
  title        = {Pith review of: Exploring Cross-Utterance Speech Contexts for Conformer-Transducer Speech Recognition Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJPA2BYX}},
  note         = {Machine review of arXiv:2508.10456}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper investigates four types of cross-utterance speech contexts modeling approaches for streaming and non-streaming Conformer-Transformer (C-T) ASR systems: i) input audio feature concatenation; ii) cross-utterance Encoder embedding concatenation; iii) cross-utterance Encoder embedding pooling projection; or iv) a novel chunk-based approach applied to C-T models for the first time. An efficient batch-training scheme is proposed for contextual C-Ts that uses spliced speech utterances within each minibatch to minimize the synchronization overhead while preserving the sequential order of cross-utterance speech contexts. Experiments are conducted on four benchmark speech datasets across three languages: the English GigaSpeech and Mandarin Wenetspeech corpora used in contextual C-T models pre-training; and the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets used in domain fine-tuning. The best performing contextual C-T systems consistently outperform their respective baselines using no cross-utterance speech contexts in pre-training and fine-tuning stages with statistically significant average word error rate (WER) or character error rate (CER) reductions up to 0.9%, 1.1%, 0.51%, and 0.98% absolute (6.0%, 5.4%, 2.0%, and 3.4% relative) on the four tasks respectively. Their performance competitiveness against Wav2vec2.0-Conformer, XLSR-128, and Whisper models highlights the potential benefit of incorporating cross-utterance speech contexts into current speech foundation models.

Figures

Figures reproduced from arXiv: 2508.10456 by Chengxi Deng, Guinan Li, Jiajun Deng, Jiawen Kang, Mengzhe Geng, Mingyu Cui, Shujie Hu, Tianzi Wang, Xie Chen, Xunying Liu, Zhaoqing Li.

Figure 1
Figure 1. Figure 1: An example of cross-utterance speech contexts (red color) and cross￾utterance textual contexts (blue color). and evaluated at the single utterance level, while longer range cross-utterance speech contexts are not fully utilized. To this end, existing researches have largely focused on incorporating two broad types of cross-utterance contexts into these systems: 1) cross-utterance textual contexts, which ha… view at source ↗
Figure 2
Figure 2. Figure 2: Example of: (a) Standard Conformer-Transducer (C-T) models operating in non-streaming or streaming mode of Section II-C, and also indicated by their use of full or triangular Encoder attention mask matrices to control input audio context (top corner of sub-figure (a)); Various C-T models modeling cross-utterance speech contexts using: (b) input audio concatenation of Section III-A; (c) cross-utterance Enco… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of data serialization during batch-mode training of contextual C-T systems without (top) or with (bottom) neighbouring utterances splicing [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: T-SNE visualizations of the context-fused encoder representations obtained using the three proposed fusion methods after the 12th Encoder transformer block: cross-utterance Encoder embedding concatenation (method (B)), cross-utterance Encoder embedding pooling projection (method (C)) and chunk-based cross-utterance Encoder embeddings (method (D)). Speakers POD1000000011 and POD1000000005 are from the GigaS… 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

100 extracted references · 68 canonical work pages · 11 internal anchors

  1. [1]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves et al. , “Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” in ICML, 2006

    2006

  2. [2]

    Hybrid ctc/attention architecture for end-to-end speech recognition,

    S. Watanabe et al. , “Hybrid ctc/attention architecture for end-to-end speech recognition,” JSTSP, 2017

    2017

  3. [3]

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

    W. Chan et al. , “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” in ICASSP, 2016

    2016

  4. [4]

    Attention is all you need,

    A. Vaswani et al. , “Attention is all you need,” in NIPS, 2017

    2017

  5. [5]

    Speech-transformer: a no-recurrence sequence-to- sequence model for speech recognition,

    L. Dong et al. , “Speech-transformer: a no-recurrence sequence-to- sequence model for speech recognition,” in ICASSP, 2018

    2018

  6. [6]

    A comparative study on transformer vs rnn in speech applications,

    S. Karita et al., “A comparative study on transformer vs rnn in speech applications,” in ASRU Workshop, 2019

    2019

  7. [7]

    Conformer: Convolution-augmented Transformer for Speech Recognition,

    A. Gulati et al. , “Conformer: Convolution-augmented Transformer for Speech Recognition,” in INTERSPEECH, 2020

    2020

  8. [8]

    Recent developments on espnet toolkit boosted by conformer,

    P. Guo et al. , “Recent developments on espnet toolkit boosted by conformer,” in ICASSP 2021 , 2021

    2021

  9. [9]

    Sequence transduction with recurrent neural networks,

    A. Graves, “Sequence transduction with recurrent neural networks,” arXiv preprint arXiv:1211.3711 , 2012

    Pith/arXiv arXiv 2012

  10. [10]

    Recurrent neural networks,

    L. R. Medsker et al. , “Recurrent neural networks,” Design and Appli- cations, 2001

    2001

  11. [11]

    Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition,

    H. Sak et al., “Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition,” arXiv preprint arXiv:1402.1128, 2014

    Pith/arXiv arXiv 2014

  12. [12]

    Efficient training of neural transducer for speech recognition,

    W. Zhou et al. , “Efficient training of neural transducer for speech recognition,” INTERSPEECH, 2022

    2022

  13. [13]

    On the limit of english conversational speech recog- nition,

    Z. T ¨uske et al. , “On the limit of english conversational speech recog- nition,” INTERSPEECH, 2021

    2021

  14. [14]

    Improving the training recipe for a robust conformer-based hybrid model,

    M. Zeineldeen et al. , “Improving the training recipe for a robust conformer-based hybrid model,” INTERSPEECH, 2022

    2022

  15. [15]

    Confidence score based conformer speaker adaptation for speech recognition,

    J. Deng et al. , “Confidence score based conformer speaker adaptation for speech recognition,” INTERSPEECH, 2022

    2022

  16. [16]

    Diagonal State Space Augmented Transformers for Speech Recognition

    G. Saon et al. , “Diagonal state space augmented transformers for speech recognition,” arXiv preprint arXiv:2302.14120 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    Recent advances in end-to-end automatic speech recog- nition,

    J. Li et al. , “Recent advances in end-to-end automatic speech recog- nition,” APSIPA Transactions on Signal and Information Processing , 2022

    2022

  18. [18]

    Improving Transformer-based Conversational ASR by Inter-Sentential Attention Mechanism

    K. Wei et al. , “Improving transformer-based conversational ASR by inter-sentential attention mechanism,” arXiv preprint arXiv:2207.00883, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  19. [19]

    Improving asr contextual biasing with guided attention,

    J. Tang et al., “Improving asr contextual biasing with guided attention,” in ICASSP, 2024

    2024

  20. [20]

    Optimizing byte-level representation for end-to-end asr,

    R. Hsiao et al. , “Optimizing byte-level representation for end-to-end asr,” in SLT, 2024

    2024

  21. [21]

    Contextual modeling for document-level asr error correction,

    J. Jiang et al. , “Contextual modeling for document-level asr error correction,” in LREC-COLING, 2024

    2024

  22. [22]

    Promptasr for contextualized asr with controllable style,

    X. Yang et al. , “Promptasr for contextualized asr with controllable style,” in ICASSP, 2024

    2024

  23. [23]

    Conversational speech recognition by learning audio- textual cross-modal contextual representation,

    K. Wei et al. , “Conversational speech recognition by learning audio- textual cross-modal contextual representation,” IEEE/ACM TASLP , 2024

    2024

  24. [24]

    Dual-mode nam: Effective top-k context injection for end-to-end asr,

    Z. Wu et al. , “Dual-mode nam: Effective top-k context injection for end-to-end asr,” in INTERSPEECH, 2023

    2023

  25. [25]

    CPPF: A contextual and post-processing-free model for automatic speech recognition

    L. Zhang et al. , “Cppf: A contextual and post-processing-free model for automatic speech recognition,” arXiv preprint arXiv:2309.07413 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    CASA-ASR: Context-Aware Speaker-Attributed ASR

    M. Shi et al. , “Casa-asr: Context-aware speaker-attributed asr,” arXiv preprint arXiv:2305.12459, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Using external off-policy speech-to-text mappings in contextual end-to-end automated speech recognition,

    D. M. Chan et al. , “Using external off-policy speech-to-text mappings in contextual end-to-end automated speech recognition,” arXiv preprint arXiv:2301.02736, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Bring dialogue-context into rnn-t for streaming asr,

    J. Hou et al. , “Bring dialogue-context into rnn-t for streaming asr,” INTERSPEECH, 2022

    work page 2022

  29. [29]

    CopyNE: Better Contextual ASR by Copying Named Entities

    S. Zhou et al. , “Copyne: Better contextual asr by copying named entities,” arXiv preprint arXiv:2305.12839 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    Improving Neural Biasing for Contextual Speech Recognition by Early Context Injection and Text Perturbation

    R. Huang et al. , “Improving neural biasing for contextual speech recognition by early context injection and text perturbation,” arXiv preprint arXiv:2407.10303, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    Training language models for long-span cross-sentence evaluation,

    K. Irie et al. , “Training language models for long-span cross-sentence evaluation,” in ASRU, 2019

    work page 2019

  32. [32]

    LSTM-LM with Long-Term History for First-Pass Decoding in Conversational Speech Recognition

    X. Chen et al. , “LSTM-LM with long-term history for first- pass decoding in conversational speech recognition,” arXiv preprint arXiv:2010.11349, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  33. [33]

    Session-level language modeling for conversational speech,

    W. Xiong et al. , “Session-level language modeling for conversational speech,” in EMNLP, 2018

    work page 2018

  34. [34]

    Transformer-xl: attentive language models beyond a fixed-length context,

    Z. Dai et al. , “Transformer-xl: attentive language models beyond a fixed-length context,” ACL, 2019

    work page 2019

  35. [35]

    End-to-end speech recognition on conversations,

    S. Kim, “End-to-end speech recognition on conversations,” Ph.D. dissertation, Carnegie Mellon University, 2019

    work page 2019

  36. [36]

    Contextualizing ASR lattice rescoring with hybrid pointer network language model,

    D.-R. Liu et al. , “Contextualizing ASR lattice rescoring with hybrid pointer network language model,” INTERSPEECH, 2020

    work page 2020

  37. [37]

    Use of contexts in language model interpolation and adaptation,

    X. Liu et al. , “Use of contexts in language model interpolation and adaptation,” CSL, 2013

    work page 2013

  38. [38]

    Longformer: the long-document transformer,

    I. Beltagy et al. , “Longformer: the long-document transformer,” arXiv preprint arXiv:2004.05150, 2020

    Pith/arXiv arXiv 2004

  39. [39]

    Transformer language models with lstm-based cross- utterance information representation,

    G. Sun et al. , “Transformer language models with lstm-based cross- utterance information representation,” in ICASSP, 2021

    work page 2021

  40. [40]

    Ctc-assisted llm-based contextual asr,

    G. Yang et al. , “Ctc-assisted llm-based contextual asr,” in SLT, 2024

    work page 2024

  41. [41]

    Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition,

    Y . Bai et al. , “Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition,” arXiv preprint arXiv:2407.04675 , 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Contextual asr error handling with llms augmentation for goal-oriented conversational ai,

    Y . Asano et al., “Contextual asr error handling with llms augmentation for goal-oriented conversational ai,” in COLING, 2025

    work page 2025

  43. [43]

    Towards asr robust spoken language understanding through in-context learning with word confusion networks,

    K. Everson et al., “Towards asr robust spoken language understanding through in-context learning with word confusion networks,” in ICASSP, 2024

    work page 2024

  44. [44]

    Contextualization of asr with llm using phonetic retrieval- based augmentation,

    Z. Lei et al., “Contextualization of asr with llm using phonetic retrieval- based augmentation,” in ICASSP, 2025

    work page 2025

  45. [45]

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

    E. Lakomkin et al. , “End-to-end speech recognition contextualization with large language models,” in ICASSP, 2024

    work page 2024

  46. [46]

    Improving domain-specific asr with llm-generated contextual descriptions,

    J. Suh et al. , “Improving domain-specific asr with llm-generated contextual descriptions,” INTERSPEECH, 2024

    work page 2024

  47. [47]

    MaLa-ASR: Multimedia-Assisted LLM-Based ASR

    G. Yang et al. , “Mala-asr: Multimedia-assisted llm-based asr,” arXiv preprint arXiv:2406.05839, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [48]

    Contextualized speech recognition: rethinking second- pass rescoring with generative large language models,

    Y . Tang et al. , “Contextualized speech recognition: rethinking second- pass rescoring with generative large language models,” in IJCAI, 2024

    work page 2024

  49. [49]

    Improving speech recognition with prompt-based contextualized asr and llm-based re-predictor,

    N. Manh Tien Anh et al. , “Improving speech recognition with prompt-based contextualized asr and llm-based re-predictor,” in IN- TERSPEECH, 2024

    work page 2024

  50. [50]

    Using large language model for end-to-end chinese asr and ner,

    Y . Li et al. , “Using large language model for end-to-end chinese asr and ner,” arXiv preprint arXiv:2401.11382 , 2024

    Pith/arXiv arXiv 2024

  51. [51]

    Compressive Transformers for Long-range Sequence Modelling,

    J. W. Rae et al., “Compressive Transformers for Long-range Sequence Modelling,” arXiv preprint arXiv:1911.05507 , 2019

    Pith/arXiv arXiv 1911

  52. [52]

    Speaker-aware Speech-transformer,

    Z. Fan et al., “Speaker-aware Speech-transformer,” in ASRU Workshop, 2019

    work page 2019

  53. [53]

    Transformer ASR with Contextual Block Process- ing,

    E. Tsunoo et al. , “Transformer ASR with Contextual Block Process- ing,” in ASRU Workshop, 2019

    work page 2019

  54. [54]

    Transformer-based long-context end-to-end speech recognition

    T. Hori et al. , “Transformer-based long-context end-to-end speech recognition.” in INTERSPEECH, 2020

    work page 2020

  55. [55]

    Advanced long-context end-to-end speech recognition using context-expanded transformers,

    T. Hori et al. , “Advanced long-context end-to-end speech recognition using context-expanded transformers,” INTERSPEECH, 2021

    work page 2021

  56. [56]

    Context-aware end-to-end asr using self-attentive embedding and tensor fusion,

    S.-Y . Chang et al. , “Context-aware end-to-end asr using self-attentive embedding and tensor fusion,” in ICASSP, 2023

    work page 2023

  57. [57]

    Longfnt: Long-form speech recognition with factorized neural transducer,

    X. Gong et al., “Longfnt: Long-form speech recognition with factorized neural transducer,” in ICASSP, 2023

    work page 2023

  58. [58]

    Advanced long-content speech recognition with factorized neural transducer,

    X. Gong et al. , “Advanced long-content speech recognition with factorized neural transducer,” IEEE/ACM TASLP, 2024

    work page 2024

  59. [59]

    Gated Embeddings in End-to-End Speech Recognition for Conversational-Context Fusion

    S. Kim et al., “Gated embeddings in end-to-end speech recognition for conversational-context fusion,” arXiv preprint arXiv:1906.11604, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  60. [60]

    Dialog-context aware end-to-end speech recognition,

    S. Kim et al. , “Dialog-context aware end-to-end speech recognition,” in SLT Workshop, 2018

    work page 2018

  61. [61]

    Improving RNN-T ASR Accuracy Using Context Audio

    A. Schwarz et al., “Improving rnn-t asr accuracy using context audio,” arXiv preprint arXiv:2011.10538 , 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  62. [62]

    Large-context automatic speech recognition based on rnn transducer,

    A. Kojima, “Large-context automatic speech recognition based on rnn transducer,” in APSIPA ASC, 2021

    work page 2021

  63. [63]

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

    H. Futami et al. , “Phoneme-aware encoding for prefix-tree-based contextual asr,” in ICASSP, 2024

    work page 2024

  64. [64]

    Deferred NAM: Low-latency Top-K Context Injection via Deferred Context Encoding for Non-Streaming ASR

    Z. Wu et al. , “Deferred nam: Low-latency top-k context injection via deferredcontext encoding for non-streaming asr,” arXiv preprint arXiv:2404.10180, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  65. [65]

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

    K. Huang et al. , “Contextualized end-to-end speech recognition with contextual phrase prediction network,” in INTERSPEECH, 2023

    work page 2023

  66. [66]

    Semi-autoregressive streaming asr with label context,

    S. Arora et al., “Semi-autoregressive streaming asr with label context,” in ICASSP, 2024

    work page 2024

  67. [67]

    Context-aware transformer transducer for speech recognition,

    F.-J. Chang et al. , “Context-aware transformer transducer for speech recognition,” in ASRU Workshop, 2021

    work page 2021

  68. [68]

    Towards effective and compact contextual representation for conformer transducer speech recognition systems,

    M. Cui et al., “Towards effective and compact contextual representation for conformer transducer speech recognition systems,” in INTER- SPEECH, 2023

    work page 2023

  69. [69]

    Wav2vec 2.0: a framework for self-supervised learning of speech representations,

    A. Baevski et al. , “Wav2vec 2.0: a framework for self-supervised learning of speech representations,” NIPS, 2020. IEEE TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING 14

    work page 2020

  70. [70]

    Wavlm: large-scale self-supervised pre-training for full stack speech processing,

    S. Chen et al., “Wavlm: large-scale self-supervised pre-training for full stack speech processing,” JSTSP, 2022

    work page 2022

  71. [71]

    Hubert: Self-supervised speech representation learn- ing by masked prediction of hidden units,

    W.-N. Hsu et al., “Hubert: Self-supervised speech representation learn- ing by masked prediction of hidden units,” TASLP, 2021

    work page 2021

  72. [72]

    data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,

    A. Baevski et al., “data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,” in ICML, 2022

    work page 2022

  73. [73]

    XLS-R: Self-supervised Cross-lingual Speech Repre- sentation Learning at Scale,

    A. Babu et al. , “XLS-R: Self-supervised Cross-lingual Speech Repre- sentation Learning at Scale,” in INTERSPEECH, 2022

    work page 2022

  74. [74]

    Data2vec: A general framework for self-supervised learning in speech, vision and language,

    A. Baevski et al., “Data2vec: A general framework for self-supervised learning in speech, vision and language,” in ICML, 2022

    work page 2022

  75. [75]

    Robust speech recognition via large-scale weak supervision,

    A. Radford et al. , “Robust speech recognition via large-scale weak supervision,” in ICML, 2023

    work page 2023

  76. [76]

    Fairseq: A fast, extensible toolkit for sequence model- ing,

    M. Ott et al. , “Fairseq: A fast, extensible toolkit for sequence model- ing,” arXiv preprint arXiv:1904.01038 , 2019

    Pith/arXiv arXiv 1904

  77. [77]

    Zipformer: A faster and better encoder for automatic speech recognition,

    Z. Yao et al. , “Zipformer: A faster and better encoder for automatic speech recognition,” arXiv preprint arXiv:2310.11230 , 2023

    Pith/arXiv arXiv 2023

  78. [78]

    ESPnet: End-to-End Speech Processing Toolkit,

    S. Watanabe et al. , “ESPnet: End-to-End Speech Processing Toolkit,” in INTERSPEECH, 2018

    work page 2018

  79. [79]

    Gigaspeech: An evolving, multi-domain asr corpus with 10k hours of transcribed audio,

    G. Chen, “Gigaspeech: An evolving, multi-domain asr corpus with 10k hours of transcribed audio,” arXiv preprint arXiv:2106.06909 , 2021

    Pith/arXiv arXiv 2021

  80. [80]

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

    B. Zhang et al. , “Wenetspeech: A 10000+ hours multi-domain man- darin corpus for speech recognition,” in ICASSP, 2022

    work page 2022

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.