pith. sign in

arxiv: 2605.22120 · v1 · pith:UTLH646Dnew · submitted 2026-05-21 · 📡 eess.AS · cs.SD

Effective User-defined Keyword Spotting with Dual-stage Matching, Multi-modal Enrollment, and Continual Adaptation

Zhiqi Ai , Han Cheng , Shiyi Mu , Xinnuo Li , Yongjin Zhou , Shugong Xu This is my paper

Pith reviewed 2026-05-22 02:40 UTC · model grok-4.3

classification 📡 eess.AS cs.SD
keywords user-defined keyword spottingdual-stage matchingmulti-modal enrollmentcontinual adaptationphoneme matcherCTC decodingon-device KWSLibriPhrase
0
0 comments X

The pith

DMA-KWS spots user-defined keywords more reliably by first locating candidates with streaming phoneme search and then verifying them with a separate matcher.

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

The paper introduces DMA-KWS to solve three practical problems in personalized wake-word detection: telling apart similar-sounding words, handling different speakers' pronunciations, and keeping the model accurate without retraining from scratch on large datasets. It uses a two-step process where a fast CTC decoder scans audio for possible matches and a second phoneme-based checker confirms or rejects them, then fuses the user's own voice recordings with text descriptions during enrollment. A lightweight update step lets the system keep improving from a small mix of real and generated examples while changing only a few hundred thousand parameters. If these pieces work as described, devices could run accurate, speaker-specific keyword spotting locally instead of sending everything to the cloud.

Core claim

DMA-KWS achieves state-of-the-art results on the LibriPhrase Hard subset with 97.85% AUC and 6.13% EER by combining a dual-stage matching pipeline (CTC streaming phoneme search followed by QbyT phoneme verification), multi-modal enrollment that merges user speech and text embeddings, and parameter-efficient continual adaptation that updates only 187k parameters using synthetic and real data. In speaker-dependent tests it outperforms text-only enrollment, showing that the added voice information and staged verification improve discrimination of confusable words while remaining suitable for on-device use.

What carries the argument

Dual-stage matching pipeline that runs CTC decoding for streaming candidate location followed by QbyT phoneme matcher for fine verification.

If this is right

  • Confusable words are distinguished more accurately because the second-stage phoneme matcher examines candidates the first stage only flagged.
  • Registered users see higher accuracy when enrollment combines their recorded speech with text embeddings rather than text alone.
  • On-device deployment remains practical because continual adaptation changes only 187k parameters.
  • Performance stays consistent across speakers with varying pronunciations due to the multi-modal and adaptation components.
  • The system can incorporate new real and synthetic data without full retraining.

Where Pith is reading between the lines

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

  • Similar staged verification might help other on-device audio tasks such as custom command recognition in noisy rooms.
  • The low parameter count for updates suggests the approach could be combined with existing wake-word engines without increasing memory footprint much.
  • If the phoneme matcher generalizes well, the framework could support open-vocabulary keyword spotting beyond a fixed set of user words.

Load-bearing premise

The CTC-based first stage will find the right candidate segments and the second-stage matcher will reliably reject confusable words even when speakers pronounce them differently.

What would settle it

On a new test set containing many phonetically similar keywords and a diverse group of speakers, measure whether the AUC falls below 97.85% or the EER rises above 6.13% when the dual-stage pipeline is replaced by a single-stage baseline.

Figures

Figures reproduced from arXiv: 2605.22120 by Han Cheng, Shiyi Mu, Shugong Xu, Xinnuo Li, Yongjin Zhou, Zhiqi Ai.

Figure 1
Figure 1. Figure 1: Schematic overview of the proposed DMA-KWS framework. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the speaker-independent user-defined KWS system with the proposed dual-stage matching architecture. The query audio is first processed [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the speaker-dependent user-defined KWS system with the proposed multi-modal enrollment architecture. The enrollment leverages the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of sample pairs with hard and easy negatives. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE visualization of challenging keywords (e.g., ”sex” vs. ”six”). [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Heatmaps of registered and query features for DMA-KWS( [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Heatmaps of wake-up scores at each (t, u) for the CTC branch and QbyT branch, representing a two-stage process: the utterance is first filtered by the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE visualization for various phonemes of DMA-KWS in the [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: DET curve comparison of different Stage-1 fine-tuning strategies on the keyword “OK Google”. Left: models using only Stage-1 fine-tuning (Table VII); [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Recall performance on the keyword “OK Google” with Persian [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Heatmaps of wake-up scores for prefix-sharing cases. prefix (e.g., ”alexa” vs. ”Alexander”), but their phoneme sequences differ12. CTC decoding helps prevent false triggers by recognizing these phonetic differences. The second case is phonetic overlap, where there is complete phonetic overlap between the words (e.g., ”Rain” vs. ”Rainbow”), causing the model to trigger the target word in the prefix region.… view at source ↗
Figure 13
Figure 13. Figure 13: Performance-Efficiency Trade-off of the cascaded DMA-KWS on HeySnips. no significant gain in accuracy. Compared to the single-stage baseline (Recall 98.06%), DMA-KWS achieves significant performance improvements with minimal additional compu￾tational overhead. VI. CONCLUSION We propose DMA-KWS, an efficient and robust framework for user-defined keyword spotting. It integrates a coarse￾to-fine dual-stage m… view at source ↗
read the original abstract

User-defined keyword spotting (KWS) is crucial for personalized voice interaction, yet existing methods face several challenges: (1) insufficient discriminability among confusable words, (2) performance inconsistency across speakers with varying pronunciations, and (3) high data cost to ensure reliable wake-word performance. In this paper, we introduce DMA-KWS, an efficient and robust framework for user-defined keyword spotting. First, it adopts a dual-stage matching pipeline: CTC decoding with streaming phoneme search to locate candidate segments, followed by QbyT with a phoneme matcher for fine-grained verification, enabling it to better distinguish confusable words. Next, multi-modal enrollment fuses user-specific speech with text embeddings to further improve accuracy for registered users. Finally, a parameter-efficient continual adaptation mechanism performs lightweight updates using synthetic and real data. Extensive experiments demonstrate the superior performance of DMA-KWS. On the LibriPhrase Hard subset, it achieves 97.85% AUC and 6.13% EER, reaching state-of-the-art performance. In speaker-dependent settings, DMA-KWS consistently outperforms text-only enrollment, demonstrating significant performance gains. Moreover, the proposed parameter-efficient fine-tuning mechanism adapts DMA-KWS with only 187k updated parameters, further enhancing KWS performance while ensuring suitability for on-device deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents DMA-KWS, a framework for user-defined keyword spotting featuring a dual-stage matching pipeline (CTC streaming phoneme search followed by QbyT phoneme matcher), multi-modal enrollment fusing speech and text embeddings, and parameter-efficient continual adaptation using synthetic and real data. It reports state-of-the-art results on the LibriPhrase Hard subset with 97.85% AUC and 6.13% EER, and superior performance over text-only enrollment in speaker-dependent scenarios with only 187k updated parameters.

Significance. Should the claims be substantiated by comprehensive ablations and statistical analysis, the work would represent a meaningful contribution to efficient, personalized keyword spotting systems suitable for on-device deployment. The dual-stage approach and multi-modal fusion could improve robustness to confusable words and speaker variations, while the adaptation mechanism addresses data efficiency concerns in real-world applications.

major comments (2)
  1. §4.2 (Ablation studies): The central claim that the dual-stage pipeline (CTC streaming search + QbyT matcher) drives the reported gains in discriminability for confusable words lacks a direct ablation that replaces the full pipeline with a single-stage QbyT matcher while holding multi-modal enrollment and continual adaptation fixed; without this, attribution of the 97.85% AUC / 6.13% EER to the dual-stage design remains unverified.
  2. §4.1 (Main results, LibriPhrase Hard): The SOTA claim of 97.85% AUC and 6.13% EER is presented without error bars, confidence intervals, or statistical significance tests relative to baselines, which is load-bearing for confirming consistent outperformance across speaker-dependent settings and data splits.
minor comments (2)
  1. Figure 3: The visualization of the dual-stage matching pipeline would benefit from explicit annotation of the candidate segment boundaries produced by CTC search.
  2. §3.3: The parameter count for the continual adaptation (187k) is stated but the breakdown of which modules are updated (e.g., which layers in the phoneme matcher) could be clarified for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to strengthen the presentation of our results and ablations.

read point-by-point responses

  1. Referee: §4.2 (Ablation studies): The central claim that the dual-stage pipeline (CTC streaming search + QbyT matcher) drives the reported gains in discriminability for confusable words lacks a direct ablation that replaces the full pipeline with a single-stage QbyT matcher while holding multi-modal enrollment and continual adaptation fixed; without this, attribution of the 97.85% AUC / 6.13% EER to the dual-stage design remains unverified.

    Authors: We agree that a direct ablation isolating the dual-stage pipeline (CTC + QbyT) while holding multi-modal enrollment and continual adaptation fixed would provide clearer attribution of gains in discriminability for confusable words. Our existing §4.2 ablations examine component contributions but do not include this exact controlled comparison. In the revised manuscript we will add this specific ablation experiment to verify the dual-stage design's role in the reported 97.85% AUC and 6.13% EER. revision: yes

  2. Referee: §4.1 (Main results, LibriPhrase Hard): The SOTA claim of 97.85% AUC and 6.13% EER is presented without error bars, confidence intervals, or statistical significance tests relative to baselines, which is load-bearing for confirming consistent outperformance across speaker-dependent settings and data splits.

    Authors: We acknowledge that the absence of error bars, confidence intervals, and statistical significance tests limits the strength of the SOTA claims. The current manuscript reports point estimates on the LibriPhrase Hard subset. In the revision we will add standard deviations across multiple runs or speaker splits, along with appropriate statistical tests (e.g., paired t-tests) against baselines to confirm consistent outperformance. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an empirical framework (DMA-KWS) with a dual-stage pipeline, multi-modal enrollment, and parameter-efficient adaptation, evaluated via standard supervised training on LibriPhrase. No equations, derivations, or self-referential definitions are described that reduce reported metrics (e.g., 97.85% AUC) to fitted inputs by construction. Performance claims rest on external benchmarks and experimental comparisons rather than self-citation chains or ansatz smuggling. This is a normal non-finding for an applied ML paper whose central results are falsifiable outside the fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework relies on standard speech-processing assumptions rather than new postulates; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Phoneme-level representations combined with CTC decoding and QbyT matching are sufficient to distinguish confusable keywords across speakers.
    This underpins the dual-stage matching claim in the abstract.

pith-pipeline@v0.9.0 · 5784 in / 1341 out tokens · 55282 ms · 2026-05-22T02:40:40.553623+00:00 · methodology

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

64 extracted references · 64 canonical work pages · 3 internal anchors

  1. [1]

    Dual data scaling for robust two-stage user-defined keyword spotting,

    Z. Ai, H. Cheng, Y . Wang, S. Mu, Y . Zhou, and S. Xu, “Dual data scaling for robust two-stage user-defined keyword spotting,” inProc. ICASSP 2026 IEEE Int. Conf. Acoust., Speech Signal Process., 2026, pp. 18 307–18 311

    work page 2026

  2. [2]

    Alexa, siri, cortana, and more: An introduction to voice assistants,

    M. B. Hoy, “Alexa, siri, cortana, and more: An introduction to voice assistants,”Med. Ref. Serv. Q., vol. 37, pp. 81 – 88, 2018

    work page 2018

  3. [3]

    Deep spoken keyword spotting: An overview,

    I. L ´opez-Espejo, Z.-H. Tan, J. H. L. Hansen, and J. Jensen, “Deep spoken keyword spotting: An overview,”IEEE Access, vol. 10, pp. 4169–4199, 2022

    work page 2022

  4. [4]

    Attention-based end-to- end models for small-footprint keyword spotting,

    C. Shan, J. Zhang, Y . Wang, and L. Xie, “Attention-based end-to- end models for small-footprint keyword spotting,” inProc. Interspeech, 2018, pp. 2037–2041

    work page 2018

  5. [5]

    Learning audio-text agreement for open-vocabulary keyword spotting,

    H.-K. Shin, H. Han, D. Kim, S.-W. Chung, and H.-G. Kang, “Learning audio-text agreement for open-vocabulary keyword spotting,” inProc. Interspeech, 2022, pp. 1871–1875

    work page 2022

  6. [6]

    Query-by-example keyword spotting system using multi-head attention and soft-triple loss,

    J. Huang, W. Gharbieh, H. S. Shim, and E. Kim, “Query-by-example keyword spotting system using multi-head attention and soft-triple loss,” inProc. ICASSP 2021 IEEE Int. Conf. Acoust., Speech Signal Process., 2021, pp. 6858–6862

    work page 2021

  7. [7]

    Auto-KWS 2021 Challenge: Task, Datasets, and Baselines,

    J. Wang, Y . He, C. Zhao, Q. Shao, W.-W. Tu, T. Ko, H.-Y . Lee, and L. Xie, “Auto-KWS 2021 Challenge: Task, Datasets, and Baselines,” in Proc. Interspeech, 2021, pp. 4244–4248

    work page 2021

  8. [8]

    PhonMatchNet: phoneme-guided zero-shot keyword spotting for user-defined keywords,

    Y .-H. Lee and N. Cho, “PhonMatchNet: phoneme-guided zero-shot keyword spotting for user-defined keywords,” inProc. Interspeech, 2023, pp. 3964–3968

    work page 2023

  9. [9]

    MM-KWS: multi-modal prompts for multilingual user-defined keyword spotting,

    Z. Ai, Z. Chen, and S. Xu, “MM-KWS: multi-modal prompts for multilingual user-defined keyword spotting,” inProc. Interspeech, 2024, pp. 2415–2419

    work page 2024

  10. [10]

    Improving acoustic based keyword spotting using lvcsr lattices,

    P. Motlicek, F. Valente, and I. Szoke, “Improving acoustic based keyword spotting using lvcsr lattices,” inProc. ICASSP 2012 IEEE Int. Conf. Acoust., Speech Signal Process., 2012, pp. 4413–4416

    work page 2012

  11. [11]

    Quantifying the value of pronunciation lexicons for keyword search in lowresource languages,

    G. Chen, S. Khudanpur, D. Povey, J. Trmal, D. Yarowsky, and O. Yilmaz, “Quantifying the value of pronunciation lexicons for keyword search in lowresource languages,” inProc. ICASSP 2013 IEEE Int. Conf. Acoust., Speech Signal Process., 2013, pp. 8560–8564

    work page 2013

  12. [12]

    Multi-task learning and weighted cross-entropy for dnn-based keyword spotting,

    S. Panchapagesan, M. Sun, A. Khare, S. Matsoukas, A. Mandal, B. Hoffmeister, and S. Vitaladevuni, “Multi-task learning and weighted cross-entropy for dnn-based keyword spotting,” inProc. Interspeech, 2016, pp. 760–764

    work page 2016

  13. [13]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” inProc. Int. Conf. Mach. Learn., 2023, pp. 28 492–28 518

    work page 2023

  14. [14]

    CaTT- KWS: a multi-stage customized keyword spotting framework based on cascaded transducer-transformer,

    Z. Yang, S. Sun, J. Li, X. Zhang, X. Wang, L. Ma, and L. Xie, “CaTT- KWS: a multi-stage customized keyword spotting framework based on cascaded transducer-transformer,” inProc. Interspeech, 2022, pp. 1681– 1685

    work page 2022

  15. [15]

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

    B. Zhang, H. Lv, P. Guo, Q. Shao, C. Yang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zenget al., “WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” inProc. ICASSP 2022 IEEE Int. Conf. Acoust., Speech Signal Process., 2022, pp. 6182–6186

    work page 2022

  16. [16]

    Streaming small-footprint keyword spotting using sequence- to-sequence models,

    Y . He, R. Prabhavalkar, K. Rao, W. Li, A. Bakhtin, and I. Mc- Graw, “Streaming small-footprint keyword spotting using sequence- to-sequence models,” inProc. 2017 IEEE Autom. Speech Recognit. Understanding Workshop, 2017, pp. 474–481

    work page 2017

  17. [17]

    DONUT: CTC-based Query-by-Example Keyword Spotting

    L. Lugosch, S. Myer, and V . S. Tomar, “DONUT: ctc-based query-by- example keyword spotting,”arXiv:1811.10736, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  18. [18]

    Streaming keyword spotting boosted by cross-layer discrimination consistency,

    Y . Xi, H. Li, X. Gu, H. Li, Y . Jiang, and K. Yu, “Streaming keyword spotting boosted by cross-layer discrimination consistency,” inProc. ICASSP 2025 IEEE Int. Conf. Acoust., Speech Signal Process., 2025, pp. 1–5

    work page 2025

  19. [19]

    TDT-KWS: fast and accurate keyword spotting using token-and-duration transducer,

    Y . Xi, H. Li, B. Yang, H. Li, H. Xu, and K. Yu, “TDT-KWS: fast and accurate keyword spotting using token-and-duration transducer,” inProc. ICASSP 2024 IEEE Int. Conf. Acoust., Speech Signal Process., 2024, pp. 11 350–11 355

    work page 2024

  20. [20]

    MFA-KWS: effective keyword spotting with multi-head frame-asynchronous decoding,

    Y . Xi, H. Li, X. Gu, Y . Jiang, and K. Yu, “MFA-KWS: effective keyword spotting with multi-head frame-asynchronous decoding,”IEEE Trans. Audio, Speech, Lang. Process., vol. 33, pp. 3014–3027, 2025

    work page 2025

  21. [21]

    Query-by-example on- device keyword spotting,

    B. Kim, M. Lee, J. Lee, Y . Kim, and K. Hwang, “Query-by-example on- device keyword spotting,” inProc. 2019 IEEE Autom. Speech Recognit. Understanding Workshop, 2019, pp. 532–538

    work page 2019

  22. [22]

    Open-vocabulary keyword spotting with audio and text embeddings,

    N. Sacchi, A. Nanchen, M. Jaggi, and M. Cernak, “Open-vocabulary keyword spotting with audio and text embeddings,” inProc. Interspeech, 2019, pp. 3362–3366

    work page 2019

  23. [23]

    Gen- eralized keyword spotting using asr embeddings,

    R. Kirandevraj, V . K. Kurmi, V . P. Namboodiri, and C. Jawahar, “Gen- eralized keyword spotting using asr embeddings,” inProc. Interspeech, 2022, pp. 126–130

    work page 2022

  24. [24]

    End-to-end asr-free keyword search from speech,

    K. Audhkhasi, A. Rosenberg, A. Sethy, B. Ramabhadran, and B. Kings- bury, “End-to-end asr-free keyword search from speech,”IEEE J. Sel. Topics Signal Process., vol. 11, no. 8, pp. 1351–1359, 2017. JOURNAL OF LATEX CLASS FILES, DECEMBER 2025 14

    work page 2017

  25. [25]

    The 2020 personalized voice trigger challenge: Open datasets, evaluation metrics, baseline system and results,

    Y . Jia, X. Wang, X. Qin, Y . Zhang, X. Wang, J. Wang, D. Zhang, and M. Li, “The 2020 personalized voice trigger challenge: Open datasets, evaluation metrics, baseline system and results,” inProc. Interspeech, 2021, pp. 4239–4243

    work page 2020

  26. [26]

    Matching latent encoding for audio-text based keyword spotting,

    K. Nishu, M. Cho, and D. Naik, “Matching latent encoding for audio-text based keyword spotting,” inProc. Interspeech, 2023, pp. 1613–1617

    work page 2023

  27. [27]

    Open- vocabulary keyword-spotting with adaptive instance normalization,

    A. Navon, A. Shamsian, N. Glazer, G. Hetz, and J. Keshet, “Open- vocabulary keyword-spotting with adaptive instance normalization,” in Proc. ICASSP 2024 IEEE Int. Conf. Acoust., Speech Signal Process., 2024, pp. 11 656–11 660

    work page 2024

  28. [28]

    Flexible keyword spotting based on homogeneous audio-text embedding,

    K. Nishu, M. Cho, P. Dixon, and D. Naik, “Flexible keyword spotting based on homogeneous audio-text embedding,” inProc. ICASSP 2024 IEEE Int. Conf. Acoust., Speech Signal Process., 2024, pp. 5050–5054

    work page 2024

  29. [29]

    Phoneme-level contrastive learning for user-defined keyword spotting with flexible enrollment,

    K. Li, H. Zhou, K. Shen, Y . Dai, and J. Du, “Phoneme-level contrastive learning for user-defined keyword spotting with flexible enrollment,” in Proc. ICASSP 2025 IEEE Int. Conf. Acoust., Speech Signal Process., 2025, pp. 1–5

    work page 2025

  30. [30]

    Contrastive learning with audio discrimination for customizable keyword spotting in continuous speech,

    Y . Xi, B. Yang, H. Li, J. Guo, and K. Yu, “Contrastive learning with audio discrimination for customizable keyword spotting in continuous speech,” inProc. ICASSP 2024 IEEE Int. Conf. Acoust., Speech Signal Process., 2024, pp. 11 666–11 670

    work page 2024

  31. [31]

    SLiCK: exploiting subsequences for length-constrained keyword spotting,

    K. Nishu, M. Cho, and D. Naik, “SLiCK: exploiting subsequences for length-constrained keyword spotting,” inProc. ICASSP 2025 IEEE Int. Conf. Acoust., Speech Signal Process., 2025, pp. 1–5

    work page 2025

  32. [32]

    Adversarial deep metric learning for cross-modal audio-text alignment in open-vocabulary keyword spotting,

    Y . Jung, Y .-H. Lee, M. Jung, J. Roh, C. W. Han, and H.-Y . Cho, “Adversarial deep metric learning for cross-modal audio-text alignment in open-vocabulary keyword spotting,” inProc. Interspeech, 2025, pp. 2645–2649

    work page 2025

  33. [33]

    Fully end-to-end streaming open-vocabulary keyword spotting with W-CTC forced alignment,

    D. Kim and J. Hwang, “Fully end-to-end streaming open-vocabulary keyword spotting with W-CTC forced alignment,” inProc. Interspeech, 2025, pp. 519–523

    work page 2025

  34. [34]

    Open vocabulary keyword spotting through transfer learning from speech synthesis,

    K. V and A. Vuppala, “Open vocabulary keyword spotting through transfer learning from speech synthesis,” inProc. Int. Conf. Signal Process. Commun., 2024, pp. 1–5

    work page 2024

  35. [35]

    CTC-aligned Audio- Text Embedding for Streaming Open-vocabulary Keyword Spotting,

    S. Jin, Y . Jung, S. Lee, J. Roh, C. Han, and H. Cho, “CTC-aligned Audio- Text Embedding for Streaming Open-vocabulary Keyword Spotting,” in Proc. Interspeech, 2024, pp. 332–336

    work page 2024

  36. [36]

    Duration-aware phone embedding upsampling for open vocabulary keyword spotting,

    R. Gundluru, N. Doppa, and S. R. M. K, “Duration-aware phone embedding upsampling for open vocabulary keyword spotting,” inProc. Natl. Conf. Commun., 2025, pp. 1–6

    work page 2025

  37. [37]

    Wav2kws: Transfer learning from speech representations for keyword spotting,

    D. Seo, H.-S. Oh, and Y . Jung, “Wav2kws: Transfer learning from speech representations for keyword spotting,”IEEE Access, vol. 9, pp. 80 682– 80 691, 2021

    work page 2021

  38. [38]

    WeKws: a production first small-footprint end-to-end keyword spotting toolkit,

    J. Wang, M. Xu, J. Hou, B. Zhang, X.-L. Zhang, L. Xie, and F. Pan, “WeKws: a production first small-footprint end-to-end keyword spotting toolkit,” inProc. ICASSP 2023 IEEE Int. Conf. Acoust., Speech Signal Process.IEEE, 2023, pp. 1–5

    work page 2023

  39. [39]

    Low-bit quantization and quantization-aware training for small-footprint keyword spotting,

    Y . Mishchenko, Y . Goren, M. Sun, C. Beauchene, S. Matsoukas, O. Rybakov, and S. N. P. Vitaladevuni, “Low-bit quantization and quantization-aware training for small-footprint keyword spotting,” in Proc. 2019 IEEE Int. Conf. Mach. Learn. Appl., 2019, pp. 706–711

    work page 2019

  40. [40]

    Convolutional neural networks for small- footprint keyword spotting,

    T. N. Sainath and C. Parada, “Convolutional neural networks for small- footprint keyword spotting,” inProc. Interspeech, 2015, pp. 1478–1482

    work page 2015

  41. [41]

    Metric learning for keyword spotting,

    J. Huh, M. Lee, H. Heo, S. Mun, and J. S. Chung, “Metric learning for keyword spotting,” inProc. 2021 IEEE Spoken Lang. Technol. Workshop, 2021, pp. 133–140

    work page 2021

  42. [42]

    Generalized end-to-end loss for speaker verification,

    L. Wan, Q. Wang, A. Papir, and I. L. Moreno, “Generalized end-to-end loss for speaker verification,” inProc. ICASSP 2018 IEEE Int. Conf. Acoust., Speech Signal Process., 2018, pp. 4879–4883

    work page 2018

  43. [43]

    The dku system description for the interspeech 2021 auto-kws challenge,

    Y . Wang, Y . Jia, M. Ma, Z. Cai, and M. Li, “The dku system description for the interspeech 2021 auto-kws challenge,”arXiv:2104.04993, 2021

    work page arXiv 2021

  44. [44]

    Deep convolutional acoustic word embeddings using word-pair side information,

    H. Kamper, W. Wang, and K. Livescu, “Deep convolutional acoustic word embeddings using word-pair side information,” inProc. ICASSP 2016 IEEE Int. Conf. Acoust., Speech Signal Process.IEEE, 2016, pp. 4950–4954

    work page 2016

  45. [45]

    Acoustic span embeddings for multilingual query-by-example search,

    Y . Hu, S. Settle, and K. Livescu, “Acoustic span embeddings for multilingual query-by-example search,” inProc. 2021 IEEE Spoken Lang. Technol. Workshop, 2021, pp. 935–942

    work page 2021

  46. [46]

    Bridging the gap between audio and text using parallel-attention for user-defined keyword spotting,

    Y . Kim, J. Jung, J. Park, B.-Y . Kim, and J. S. Chung, “Bridging the gap between audio and text using parallel-attention for user-defined keyword spotting,”IEEE Signal Process. Lett., vol. 31, pp. 2100–2104, 2024

    work page 2024

  47. [47]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    V . Sanh, L. Debut, J. Chaumond, and T. Wolf, “DistilBERT, a distilled version of bert: smaller, faster, cheaper and lighter,”arXiv:1910.01108, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  48. [48]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 12 449–12 460

    work page 2020

  49. [49]

    LLM-Synth4KWS: scalable automatic generation and synthesis of confusable data for custom keyword spotting,

    P. Zhu, Q. Wang, D. Agarwal, and K. Partridge, “LLM-Synth4KWS: scalable automatic generation and synthesis of confusable data for custom keyword spotting,” inProc. Interspeech, 2025, pp. 2675–2679

    work page 2025

  50. [50]

    Utilizing tts synthesized data for efficient development of keyword spotting model,

    H. J. Park, D. Agarwal, N. Chen, R. Sun, K. Partridge, J. Chen, H. Zhang, P. Zhu, J. W. Bartel, K. Kastner, Y . Wang, A. Rosenberg, and Q. Wang, “Utilizing tts synthesized data for efficient development of keyword spotting model,” inProc. ISCA SynData4GenAI Workshop, 2024, pp. 16–20

    work page 2024

  51. [51]

    Fully unsupervised training of few-shot keyword spotting,

    D. Lee, M. Kim, S. H. Mun, M. H. Han, and N. S. Kim, “Fully unsupervised training of few-shot keyword spotting,” inProc. 2023 IEEE Spoken Lang. Technol. Workshop, 2023, pp. 266–272

    work page 2023

  52. [52]

    Synth4Kws: synthesized speech for user defined keyword spotting in low resource environments,

    P. Zhu, D. Agarwal, J. W. Bartel, K. Partridge, H. J. Park, and Q. Wang, “Synth4Kws: synthesized speech for user defined keyword spotting in low resource environments,” inProc. ISCA SynData4GenAI Workshop, 2024, pp. 11–15

    work page 2024

  53. [53]

    Conformer: Convolution- augmented Transformer for Speech Recognition,

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

    work page 2020

  54. [54]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inProc. Int. Conf. Learn. Represent., 2022. [Online]. Available: https://openreview.net/forum?id=nZeVKeeFYf9

    work page 2022

  55. [55]

    F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,” inProc. ACL, 2025, pp. 6255–6271

    work page 2025

  56. [56]

    LibriSpeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “LibriSpeech: An ASR corpus based on public domain audio books,” inProc. ICASSP 2015 IEEE Int. Conf. Acoust., Speech Signal Process., 2015, pp. 5206– 5210

    work page 2015

  57. [57]

    GigaSpeech: an evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,

    G. Chen, S. Chai, G.-B. Wang, J. Du, W.-Q. Zhang, C. Weng, D. Su, D. Povey, J. Trmal, J. Zhang, M. Jin, S. Khudanpur, S. Watanabe, S. Zhao, W. Zou, X. Li, X. Yao, Y . Wang, Z. You, and Z. Yan, “GigaSpeech: an evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,” inProc. Interspeech, 2021, pp. 3670–3674

    work page 2021

  58. [58]

    Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition

    P. Warden, “Speech Commands: A dataset for limited-vocabulary speech recognition,”arXiv:1804.03209, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  59. [59]

    AudioMNIST: Exploring explainable artificial intelli- gence for audio analysis on a simple benchmark,

    S. Becker, J. Vielhaben, M. Ackermann, K.-R. M ¨uller, S. Lapuschkin, and W. Samek, “AudioMNIST: Exploring explainable artificial intelli- gence for audio analysis on a simple benchmark,”J. Franklin Inst., vol. 361, no. 1, pp. 418–428, 2024

    work page 2024

  60. [60]

    Efficient keyword spotting using dilated convolutions and gating,

    A. Coucke, M. Chlieh, T. Gisselbrecht, D. Leroy, M. Poumeyrol, and T. Lavril, “Efficient keyword spotting using dilated convolutions and gating,” inProc. ICASSP 2018 IEEE Int. Conf. Acoust., Speech Signal Process., 2018, pp. 6351–6355

    work page 2018

  61. [61]

    A multi purpose and large scale speech corpus in persian and english for speaker and speech recognition: The deepmine database,

    H. Zeinali, L. Burget, and J. H. ˇCernock´y, “A multi purpose and large scale speech corpus in persian and english for speaker and speech recognition: The deepmine database,” inProc. 2019 IEEE Autom. Speech Recognit. Understanding Workshop, 2019, pp. 397–402

    work page 2019

  62. [62]

    SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,” inProc. Interspeech, 2019, pp. 2613– 2617

    work page 2019

  63. [63]

    WavLM: Large-scale self-supervised pre- training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “WavLM: Large-scale self-supervised pre- training for full stack speech processing,”IEEE J. Sel. Topics Signal Process., vol. 16, no. 6, pp. 1505–1518, 2022

    work page 2022

  64. [64]

    Re-weighted interval loss for handling data imbalance problem of end- to-end keyword spotting,

    K. Zhang, Z. Wu, D. Yuan, J. Luan, J. Jia, H. Meng, and B. Song, “Re-weighted interval loss for handling data imbalance problem of end- to-end keyword spotting,” inProc. Interspeech, 2020, pp. 2567–2571

    work page 2020