Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Unified Semi-Supervised Pipeline for Automatic Speech Recognition

T0 review · 3 major / 6 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 A fully open-source semi-supervised ASR pipeline, built around the TopIPL pseudo-labeling algorithm, reports relative word-error-rate reductions of 18–40% for Portuguese, 5–16% for Armenian, and 2–8% for Spanish, and outperforms prior…

desk verdict Useful open-source pipeline and a plausible pseudo-labeling variant, but the Portuguese MLS evaluation looks contaminated and the headline gains need a dedup check before I'd trust them. read the letter →

arxiv 2506.07659 v1 pith:3BTEVUEP submitted 2025-06-09 eess.AS

Nune Tadevosyan , Nikolay Karpov , Andrei Andrusenko , Vitaly Lavrukhin , Ante Jukic This is my paper
classification eess.AS
keywords speechrecognitionpseudo-labelingsemi-supervisedlearningdatapipelinePortugueseSpanishArmenianworderrorrate
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 tries to establish that a complete, open-source, semi-supervised speech-recognition pipeline—covering unlabeled data collection from public audio, cleaning, pseudo-labeling, and model training—can deliver consistent accuracy gains across languages with very different amounts of labeled data. Its proposed pseudo-labeling algorithm, TopIPL, waits until a model has fully converged on labeled data before generating transcriptions for unlabeled audio, then refreshes those pseudo-labels from a cache during training and refines them with a teacher built from the student's best checkpoints. The reported results are relative word-error-rate reductions of 18–40% for Portuguese with 80 hours of labeled data, 5–16% for Armenian with 57 hours, and 2–8% for Spanish with 1400 hours, with the best LibriSpeech-100/860 test-clean WER of 5.13 for the RNNT model, ahead of the EMA baseline. A sympathetic reader would care because it suggests low-resource languages can get large accuracy gains from messy, publicly available speech without manual curation, and the whole recipe is released for reuse.

What carries the argument

The load-bearing mechanism is TopIPL's staged pseudo-labeling schedule with a dynamic cache and a checkpoint-averaged teacher. In the first stage, a student model trains only on labeled data until full convergence; only then are pseudo-labels generated for the entire unlabeled pool in one pass and stored in a cache. In the second stage, the student trains on labeled data plus the cache, and at each epoch's end the cache is refreshed with new pseudo-labels with probability p_cache = 0.2. In the third stage, the teacher is initialized from the student and updated by averaging the student's top-N checkpoints by validation WER, and the teacher regenerates pseudo-labels for the cache. The surrounding pipeline converts audio to 16 kHz WAV, uses a language-identification model with segment-based voting to keep only the target language, applies a forced aligner to create 1–20 second segments with punctuation boundaries, and optionally filters by character rate. Pseudo-labels are generated with the faster CTC (connectionist temporal classification) head, while the reported gains are measured on the RNNT (recurrent neural network transducer) head.

What would settle it

Run TopIPL on the Portuguese 80-hour setup with p_cache set to 0 so pseudo-labels are generated only once, and compare WER on mTEDx and CORAA against the p_cache = 0.2 configuration; if the two are statistically indistinguishable, the dynamic cache refresh—a stated mechanism—is not responsible for the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that delaying pseudo-labeling until the labeled-data model has converged, then alternating between training on labeled plus pseudo-labeled data and refreshing a pseudo-label cache with probability 0.2, produces a more accurate and stable student than the usual EMA-style teacher averaging. The teacher is not a running average of all weights; it is formed by averaging the top-N student checkpoints selected by validation WER, with top-3 used in the main experiments, which the authors argue captures the most informative model states. The paper further claims this works on real-world noisy audio without explicit pseudo-label filtering: character-rate filtering between 5 and 21 characters per second did not consistently help, and the unfiltered noisy data often trained better models. On LibriSpeech-100/860, TopIPL reaches 5.13% test-clean WER with the RNNT head, beating the EMA comparison at 5.58% and other pseudo-labeling variants.

Load-bearing premise

The method's gains rest on the baseline model's pseudo-labels for new, noisy unlabeled audio being accurate enough that training on them improves the model instead of reinforcing systematic transcription errors.

Editorial extensions

If this is right

  • If TopIPL's reported gains hold, a team with as little as 50–80 hours of transcribed speech can cut word error rates by up to 40 percent on difficult test sets by adding about 2,000 hours of permissively licensed web audio through the released pipeline.
  • The LibriSpeech-100/860 results imply that delayed pseudo-labeling with checkpoint-averaged teachers is a stronger default than EMA-style momentum teachers for semi-supervised ASR.
  • Because character-rate filtering did not consistently improve WER and sometimes worsened it, large noisy unlabeled corpora can be used without expensive cleaning, lowering the cost of scaling to new languages.
  • The open-source release of the full pipeline means the data-collection, filtering, and training stages can be reused or modified for languages beyond the three tested.
  • There is an optimal amount of unlabeled data: 2,000 hours helped Spanish and Portuguese most, while 2,500 hours of Portuguese degraded accuracy, so practitioners should tune unlabeled pool size rather than assume more is better.

Reading between the lines

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

  • The authors do not run an ablation of the cache-refresh probability p_cache. A reader could test whether p_cache = 0, with pseudo-labels generated only once, performs as well as p_cache = 0.2; if it does, the dynamic refresh stage is not the active ingredient.
  • The 5–21 character-rate filter is implicitly calibrated on the three tested languages; languages with very different orthographies may need different thresholds, and the paper gives no guidance on how to set them.
  • Because pseudo-labels are generated with the fast CTC head while gains appear on the RNNT head, the approach may transfer to streaming or on-device transducers, where RNNT is the deployment head.
  • The pipeline depends on the language-identification and forced-alignment stages being reliable for a new language, but the paper does not evaluate how sensitive TopIPL's gains are to errors in those stages.
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 / 6 minor

Summary. This paper proposes a fully open-source semi-supervised ASR pipeline, covering unlabeled data collection from Yodas and YouTube Commons, language filtering, segmentation, and character-rate filtering, together with a new pseudo-labeling algorithm called TopIPL. TopIPL first trains a model on labeled data to convergence, generates pseudo-labels for a cache of unlabeled data, then alternates between training on labeled and pseudo-labeled data, periodically refreshing the cache with a probability p_cache, and finally uses a teacher model formed by averaging the top-N student checkpoints by validation performance. The method is evaluated on Portuguese (MCV 80h labeled, MLS plus web data as unlabeled), Armenian (MCV+Fleurs 57h labeled, 145h unlabeled), Spanish (1400h labeled, 10k h unlabeled), and on LibriSpeech-100/860. The reported results show consistent relative WER improvements over labeled-only baselines across all languages (18-40% for Portuguese, 5-16% for Armenian, 2-8% for Spanish) and state-of-the-art-like results on LibriSpeech-100/860 compared with previous semi-supervised approaches.

Significance. If the reported results are correct and the data hygiene concerns are resolved, this contribution is practically significant: it provides an end-to-end, reproducible recipe for building ASR models from noisy web-scale unlabeled speech, and it gives evidence that a relatively simple pseudo-labeling scheme with a dynamic cache and top-N checkpoint averaging can outperform established techniques such as EMA. The LibriSpeech ablation is a useful controlled comparison among pseudo-labeling strategies. However, the external-validity threat from potential train/test overlap in the Portuguese setup is a serious caveat, and the absence of variance estimates and hyperparameter sensitivity analyses limits the strength of the empirical claims.

major comments (3)
  1. [§3.1.2, §2.1, Table 2] The Portuguese experimental setup has a potentially load-bearing data leakage risk. Section 3.1.2 states that MLS (160h) is treated as an unlabeled source, and Table 2 reports WER on MLS as an evaluation set. The paper does not state that the train/test splits of MLS are kept disjoint. Additionally, Section 2.1 sources unlabeled data from Yodas and YouTube Commons; mTEDx consists of TEDx talks that are widely mirrored on YouTube, and CORAA contains spontaneous recordings whose provenance is not documented. No audio-level or transcript-level deduplication between the unlabeled pool and the mTEDx/CORAA test sets is reported. If overlap occurs, the headline 18-40% relative WER improvements on Portuguese (Section 4.2) would reflect memorization rather than generalization. The authors should either provide a deduplication procedure, or explicitly state and verify disjointness of all unlabeled training data from all evaluation sets, or qualify the claims accordingly.
  2. [§2.5, §4.1, §4.2] The method's free parameters p_cache=0.2 and top-N=3 are selected through 'extensive experimentation' (Section 2.5) but no ablations or sensitivity analyses are reported. The character-rate thresholds CR_min=5 and CR_max=21 (Section 2.4) are also presented as fixed heuristics without supporting experiments. Furthermore, no standard deviations or multiple-seed results are given for any WER number. Given that the optimal amount of unlabeled data varies across settings (e.g., 2k h vs. 2.5k h for Portuguese in Table 2), the robustness of the central claim to these choices is unclear. Please provide ablation studies over p_cache and top-N, and report results over at least a few seeds for the main reported configurations.
  3. [§4.3, Table 3] The LibriSpeech-100/860 comparison lacks crucial details for reproducibility and fair comparison. For TopIPL, the top-N checkpoints are 'selected by validation performance' (Section 2.5), but the text does not specify which validation set was used for LibriSpeech experiments—if it was test-clean/test-other, the reported WERs would be optimistically biased. For the EMA baseline, no implementation details are given (decay factor, update schedule, whether it follows the same training stages and epochs). The claim that TopIPL 'outperforms previous works' (Section 4.3) is based on this comparison, so these details are essential. Please clarify the validation protocol for LibriSpeech and describe the EMA baseline precisely.
minor comments (6)
  1. [Table 2] The caption contains a typo: 'Portuges e' should be 'Portuguese'.
  2. [§4.1] The sentence 'Our analysis show that adding 2k hours...' should read 'Our analysis shows...'.
  3. [§4.2] The phrases 'filtered 2.4 datasets' and 'the filtered 2.4 data' should refer to 'Section 2.4' rather than '2.4'.
  4. [§2] The sentence 'We follow a structured pipeline 2' should reference 'Figure 1' instead of 'pipeline 2'.
  5. [Algorithm 1] The loop labels 'For n epochs do', 'For m epochs do', and 'Form epochs do' are somewhat ambiguous about which training stage each refers to; consider clarifying the stage boundaries and indentation.
  6. [References] Reference [17] (YouTube Commons) appears non-standard and could be formatted more completely; also consider citing the specific mTEDx and CORAA versions used.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: WER gains are empirical measurements on external held-out test sets; the only mild issue is selection-on-validation, which is not circular.

full rationale

The paper's central claims (18–40% relative WER gains for Portuguese, 5–16% for Armenian, 2–8% for Spanish, and 5.13 test-clean WER on LibriSpeech) are empirical measurements of trained models against held-out test sets (mTEDx, CORAA, Spanish test partitions, LibriSpeech test-clean/test-other), not derivations that reduce to their inputs. Self-training is inherently self-referential — Algorithm 1 generates pseudo-labels with the student model S and builds a teacher T from S's own top-N checkpoints — but the evaluation is external to those pseudo-labels, so the claim that pseudo-labeling improves WER is not forced by construction. The reference list contains no works authored by the present paper's authors, so no self-citation is load-bearing; no uniqueness theorem or ansatz is imported via citation; and TopIPL is explicitly compared against EMA and iterative pseudo-labeling baselines rather than being a renamed known result. The only mild issue is selection-on-validation: the teacher is formed from 'top-3 checkpoints of the student model based on its performance on the MCV validation set' (Section 4.1), hyperparameters were 'tuned based on validation performance across all datasets using WER' (Section 3.1.3), and p_cache=0.2 was chosen 'through extensive experimentation' (Section 2.5). Reporting MCV test WER after tuning on MCV validation, and reporting gains on the same benchmark family used for tuning, can introduce optimism bias, but the test sets are disjoint from the tuning sets, so the reported numbers are not statistically forced. The skeptic's concern about possible overlap between the YouTube-sourced unlabeled pool (Yodas/YouTube Commons) and the mTEDx/CORAA evaluation sets is an external-validity threat, not circularity: Section 3.1.2 states only that those sets are 'reserved for evaluation only' because of their CC BY-NC-ND license and reports no deduplication, but even if leakage occurred it would invalidate the measurement rather than making it definitionally equivalent to its inputs. Verdict: no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on several domain assumptions about unlabeled data quality and pseudo-label accuracy, plus four hand-tuned hyperparameters (p_cache, top-N, CR bounds). No invented entities are introduced.

free parameters (4)
  • p_cache = 0.2
    Cache update probability, tuned through extensive experimentation (Section 2.5), no ablation shown.
  • top_N = 3
    Number of student checkpoints averaged into the teacher, chosen for all experiments without ablation.
  • CR_min = 5
    Lower character-rate threshold for filtering unlabeled audio (Section 2.4), set by hand.
  • CR_max = 21
    Upper character-rate threshold (Section 2.4), set by hand.
assumptions (3)
  • domain assumption Publicly available speech under Creative Commons licenses provides enough usable unlabeled data for any target language.
    The whole framework rests on this; stated in the abstract and Section 2.1.
  • domain assumption AmberNet language identification and NeMo Forced Aligner segmentation errors are low enough that the resulting unlabeled segments are suitable for pseudo-labeling.
    Sections 2.2 and 2.3; if the language ID or segmentation is wrong, pseudo-labels will not match the target language.
  • domain assumption Baseline ASR pseudo-labels are sufficiently accurate for self-training to help.
    Algorithm 1 lines 4 and 8: pseudo-labels come from the student model trained only on labeled data; large errors could propagate through cache updates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unified Semi-Supervised Pipeline for Automatic Speech Recognition." pith.science (2026). https://pith.science/paper/3BTEVUEP

@misc{pith2026250607659,
  author       = {Pith},
  title        = {Pith review of: Unified Semi-Supervised Pipeline for Automatic Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BTEVUEP}},
  note         = {Machine review of arXiv:2506.07659}
}
read the original abstract

Automatic Speech Recognition has been a longstanding research area, with substantial efforts dedicated to integrating semi-supervised learning due to the scarcity of labeled datasets. However, most prior work has focused on improving learning algorithms using existing datasets, without providing a complete public framework for large-scale semi-supervised training across new datasets or languages. In this work, we introduce a fully open-source semi-supervised training framework encompassing the entire pipeline: from unlabeled data collection to pseudo-labeling and model training. Our approach enables scalable dataset creation for any language using publicly available speech data under Creative Commons licenses. We also propose a novel pseudo-labeling algorithm, TopIPL, and evaluate it in both low-resource (Portuguese, Armenian) and high-resource (Spanish) settings. Notably, TopIPL achieves relative WER improvements of 18-40% for Portuguese, 5-16% for Armenian, and 2-8% for Spanish.

Figures

Figures reproduced from arXiv: 2506.07659 by the authors.

Figure 1
Figure 1. Proposed semi-supervised pipeline for training ASR models. • Open-source release of the data collection pipeline and Top￾IPL, promoting reproducibility and accessibility. 2. Method We follow a structured pipeline2 to process large-scale speech data, ensuring clean and language-specific training samples. Our approach can be applicable to any language and integrates NeMo tools, including the Speech Data Processor (SDP… view at source ↗
Figure 2
Figure 2. WER improvements across languages and datasets. The transparent part represents the baseline performance, while the opaque part is the performance after applying TopIPL [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

  1. [1]

    Semi-supervised learning miti- gates this challenge by leveraging large-scale unlabeled speech data

    Introduction Training of Automatic Speech Recognition (ASR) models re- quires large amounts of labeled data, which is often expensive and time-consuming to obtain. Semi-supervised learning miti- gates this challenge by leveraging large-scale unlabeled speech data. However, acquiring such data presents its own difficulties, since data collected from real-w...

    work page

  2. [2]

    Our approach can be applicable to any language and integrates NeMo tools, including the Speech Data Processor (SDP) 3, to automate data collection and preprocessing

    Method We follow a structured pipeline 2 to process large-scale speech data, ensuring clean and language-specific training samples. Our approach can be applicable to any language and integrates NeMo tools, including the Speech Data Processor (SDP) 3, to automate data collection and preprocessing. The main stages of the proposed pipeline are illustrated in...

    work page

  3. [3]

    Experimental Setup For our experiments, we utilize the NVIDIA Fast Conformer encoder [22], which is 2.8 times faster than the standard Con- former encoder [23]. To accelerate model convergence, we employ two decoding heads: Connectionist Temporal Classi- fication (CTC) [24] and Recurrent Neural Network Transducer (RNNT) [25], with the model’s loss functio...

    work page arXiv

  4. [4]

    The most significant improvement is observed in Portuguese, where relative WER reduction reaches up to 40%

    Results In Figure 2, we present the performance of models trained us- ing the TopIPL algorithm across three different languages. The most significant improvement is observed in Portuguese, where relative WER reduction reaches up to 40%. Here, labeled data is scarce, but a sufficient amount of unlabeled data is available, allowing the model to leverage sel...

    work page arXiv

  5. [5]

    We evaluate TopIPL across three distinct setups: low-resource Portuguese and Armenian, and a high-resource Spanish observing consistent relative WER reductions in all cases

    Conclusion In this work, we introduce an open-source data preprocessing pipeline alongside a novel pseudo-labeling algorithm, TopIPL, for semi-supervised ASR training. We evaluate TopIPL across three distinct setups: low-resource Portuguese and Armenian, and a high-resource Spanish observing consistent relative WER reductions in all cases. Our results sho...

    work page

  6. [6]

    Momentum pseudo-labeling for semi-supervised speech recognition,

    Y . Higuchi, N. Moritz, J. L. Roux, and T. Hori, “Momentum pseudo-labeling for semi-supervised speech recognition,” inIn- terspeech 2021, 2021, pp. 726–730

    work page 2021

  7. [7]

    End-to-end ASR: from su- pervised to semi-supervised learning with modern architectures,

    G. Synnaeve, Q. Xu, J. Kahnet al., “End-to-end ASR: from su- pervised to semi-supervised learning with modern architectures,” inICML 2020 Workshop on Self-supervision in Audio and Speech, 2020

    work page 2020

  8. [8]

    Self-training for end-to- end speech recognition,

    J. Kahn, A. Lee, and A. Hannun, “Self-training for end-to- end speech recognition,” inICASSP 2020-2020 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 7084–7088

    2020

Show all 40 references
  1. [9]

    Iterative pseudo-labeling for speech recognition,

    Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert, “Iterative pseudo-labeling for speech recognition,” inInterspeech 2020, 2020, pp. 1006–1010

    2020

  2. [10]

    Continuous pseudo-labeling from the start,

    D. Berrebbi, R. Collobert, S. Bengio, N. Jaitly, and T. Likhoma- nenko, “Continuous pseudo-labeling from the start,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  3. [11]

    slimipl: Language-model-free iterative pseudo-labeling,

    T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert, “slimipl: Language-model-free iterative pseudo-labeling,” inIn- terspeech 2021, 2021, pp. 741–745

    2021

  4. [12]

    Censer: Curriculum semi-supervised learning for speech recog- nition based on self-supervised pre-training,

    B. Zhang, S. Cao, X. Xhang, Y . Zhang, L. Ma, and T. Shinozaki, “Censer: Curriculum semi-supervised learning for speech recog- nition based on self-supervised pre-training,” inInterspeech 2022, 2022, pp. 2653–2657

    2022

  5. [13]

    Kaizen: Continuously improving teacher using exponential moving average for semi- supervised speech recognition,

    V . Manohar, T. Likhomanenko, Q. Xu, W.-N. Hsu, R. Collobert, Y . Saraf, G. Zweig, and A. Mohamed, “Kaizen: Continuously improving teacher using exponential moving average for semi- supervised speech recognition,” in2021 IEEE Automatic Speech Recognition and Understanding Work...

    2021

  6. [14]

    Momentum pseudo-labeling: Semi-supervised asr with continuously improv- ing pseudo-labels,

    Y . Higuchi, N. Moritz, J. Le Roux, and T. Hori, “Momentum pseudo-labeling: Semi-supervised asr with continuously improv- ing pseudo-labels,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1424–1438, 2022

    2022

  7. [15]

    Despite these advancements, several shortcomings remain

    have been explored to stabilize training, reducing overfit- ting risks and ensuring more robust learning. Despite these advancements, several shortcomings remain. Firstly, many state-of-the-art approaches are not open-source, limiting reproducibility and slowing research progr...

    2025 arXiv

  8. [16]

    Advancing momentum pseudo-labeling with conformer and initialization strategy,

    ——, “Advancing momentum pseudo-labeling with conformer and initialization strategy,” inICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7672–7676

    2022

  9. [17]

    In- termpl: Momentum pseudo-labeling with intermediate ctc loss,

    Y . Higuchi, T. Ogawa, T. Kobayashi, and S. Watanabe, “In- termpl: Momentum pseudo-labeling with intermediate ctc loss,” inICASSP 2023-2023 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  10. [18]

    Improved noisy student training for automatic speech recognition,

    D. S. Park, Y . Zhang, Y . Jia, W. Han, C.-C. Chiu, B. Li, Y . Wu, and Q. V . Le, “Improved noisy student training for automatic speech recognition,” inInterspeech 2020, 2020, pp. 2817–2821

    2020

  11. [19]

    Al- ternative pseudo-labeling for semi-supervised automatic speech recognition,

    H. Zhu, D. Gao, G. Cheng, D. Povey, P. Zhang, and Y . Yan, “Al- ternative pseudo-labeling for semi-supervised automatic speech recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  12. [20]

    Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,

    S. Khurana, N. Moritz, T. Hori, and J. Le Roux, “Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,” inICASSP 2021-2021 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6553–6557

    2021

  13. [21]

    Improving pseudo-label training for end-to-end speech recognition using gradient mask,

    S. Ling, C. Shen, M. Cai, and Z. Ma, “Improving pseudo-label training for end-to-end speech recognition using gradient mask,” inICASSP 2022-2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8397–8401

    2022

  14. [22]

    Yodas: Youtube-oriented dataset for audio and speech,

    X. Li, S. Takamichi, T. Saeki, W. Chen, S. Shiota, and S. Watan- abe, “Yodas: Youtube-oriented dataset for audio and speech,” 2024

    2024

  15. [23]

    Youtube commons dataset,

    PleIAs, “Youtube commons dataset,” n.d. [Online]. Available: https://huggingface.co/datasets/PleIAs/YouTube-Commons

  16. [24]

    Creative commons attribution 4.0 in- ternational public license,

    C. Commons, “Creative commons attribution 4.0 in- ternational public license,” 2013. [Online]. Available: https://creativecommons.org/licenses/by/4.0/

    2013

  17. [25]

    Ambernet: A compact end-to-end model for spoken language identification,

    F. Jia, N. R. Koluguri, J. Balam, and B. Ginsburg, “Ambernet: A compact end-to-end model for spoken language identification,” arXiv preprint arXiv:2210.15781, 2022

    2022 arXiv

  18. [26]

    Langid ambernet checkpoints,

    NVIDIA, “Langid ambernet checkpoints,” 2023, pre- trained model available on NVIDIA NGC. [Online]. Avail- able: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/ models/langid_ambernet

    2023

  19. [27]

    Nemo forced aligner,

    ——, “Nemo forced aligner,” GitHub repository. [Online]. Avail- able: https://github.com/NVIDIA/NeMo/tree/main/tools/nemo_ forced_aligner

  20. [28]

    Fast conformer with linearly scalable attention for efficient speech recognition,

    D. Rekesh, N. R. Koluguri, S. Kriman, S. Majumdar, V . Noroozi, H. Huang, O. Hrinchuk, K. Puvvada, A. Kumar, J. Balamet al., “Fast conformer with linearly scalable attention for efficient speech recognition,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop ...

    2023

  21. [29]

    Conformer: Convolution-augmented transformer for speech recognition,

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

    2020

  22. [30]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” inProceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

    2006

  23. [31]

    Exploring architectures, data and units for streaming end-to-end speech recognition with rnn-transducer,

    K. Rao, H. Sak, and R. Prabhavalkar, “Exploring architectures, data and units for streaming end-to-end speech recognition with rnn-transducer,” in2017 IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2017, pp. 193–199

    2017

  24. [32]

    Stateful conformer with cache-based inference for streaming au- tomatic speech recognition,

    V . Noroozi, S. Majumdar, A. Kumar, J. Balam, and B. Ginsburg, “Stateful conformer with cache-based inference for streaming au- tomatic speech recognition,” inICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. ...

    2024

  25. [33]

    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,” inInterspeech 2019, 2019, pp. 2613–2617

    2019

  26. [34]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,”arXiv preprint arXiv:1912.06670, 2019

    1912 arXiv

  27. [35]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” inInter- speech 2020, 2020, pp. 2757–2761

    2020

  28. [36]

    V oxpopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation,

    C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haz- iza, M. Williamson, J. Pino, and E. Dupoux, “V oxpopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation,”arXiv preprint arXiv:2101.00390, 2021

    2021 arXiv

  29. [37]

    English hybrid fastconformer model,

    NVIDIA, “English hybrid fastconformer model,” https: //huggingface.co/nvidia/stt_en_fastconformer_ctc_large, 2023

    2023

  30. [38]

    The multilingual tedx corpus for speech recognition and translation,

    E. Salesky, M. Wiesner, J. Bremerman, R. Cattoni, M. Negri, M. Turchi, D. W. Oard, and M. Post, “The multilingual tedx corpus for speech recognition and translation,”arXiv preprint arXiv:2102.01757, 2021

    2021 arXiv

  31. [39]

    Coraa: a large corpus of spontaneous and prepared speech manually validated for speech recognition in brazilian portuguese,

    A. C. Junior, E. Casanova, A. Soareset al., “Coraa: a large corpus of spontaneous and prepared speech manually validated for speech recognition in brazilian portuguese,”arXiv preprint arXiv:2110.15731, 2021

    2021 arXiv

  32. [40]

    Fleurs: Few-shot learning evaluation of universal representations of speech,

    A. Conneau, M. Ma, S. Khanuja, Y . Zhang, V . Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna, “Fleurs: Few-shot learning evaluation of universal representations of speech,”2022 IEEE Spoken Language Technology Workshop (SLT), pp. 798–805, 2022

    2022

Pith tools

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