Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Complexity boosted adaptive training for better low resource ASR performance

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Two-stage adaptive training cuts low-resource ASR errors by up to 14 percent by tailoring augmentation and intermediate CTC loss to sample difficulty.

desk verdict Plausible training-only recipe for low-resource ASR, but the two-stage schedule confound means the adaptive policy may not be doing the work. read the letter →

arxiv 2412.00877 v1 pith:PSOFMZP3 submitted 2024-12-01 cs.SD eess.AS

Hongxuan Lu , Shenjian Wang , Biao Li This is my paper · ORCID
classification cs.SDeess.AS
keywords low-resourceASRdataaugmentationSpecAugmentintermediateCTCadaptivetrainingsamplecomplexityConformertwo-stage
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 claims that speech recognition models underperform when augmentation and auxiliary losses are applied at fixed strength to every sample. It proposes a two-stage, complexity-boosted adaptive (CBA) training scheme that estimates how hard each utterance is from its min-max normalized loss within the mini-batch and uses that estimate to set both the amount of SpecAugment masking and the weight of the intermediate CTC loss. In stage one the model trains without adaptation; in stage two a new policy, MinMax-IBF, adapts both terms. The reported result is a 13.4% and 14.1% relative WER reduction on LibriSpeech 100h test-clean and test-other over the fixed-policy Conformer baseline, plus a 6.3% relative CER reduction on AISHELL-1, with no change to the model at inference time. A reader should care because the improvement comes from a training-only recipe and could transfer to other CTC-based ASR systems.

What carries the argument

MinMax-IBF, a sample-adaptive policy built on the incomplete $\beta$ function. For each sample $i$ in a mini-batch, complexity is $x_i = (L_i - L_{\min}) / (L_{\max} - L_{\min})$; the augmentation intensity is $f_{\mathrm{DA}} = 1 - \mathrm{IBF}(x_i)$, linearly mapped to the number of time/frequency masks, and the intermediate-CTC regularization weight is $f_{\mathrm{CTC}} = (1/B)\sum_i (1 - \mathrm{IBF}(x_i))$. The incomplete $\beta$ function maps the normalized loss to a smooth policy value, while the min-max normalization distinguishes samples with equal rank but different loss magnitudes. This carries the paper's argument because the same complexity estimate controls two separate training mechanisms, and the adaptive training is only switched on in the second stage.

What would settle it

Record the within-batch loss spread during stage 2 on LibriSpeech 100h; if the median relative spread $(L_{\max} - L_{\min})/L_{\max}$ across batches is below about 0.05, then $x_i$ collapses toward 0 or 1 and the policy assigns near-extreme intensities to nearly all samples, showing the complexity measure is not what drives the reported gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that sample complexity, measured as the position of a sample's loss between the batch minimum and maximum, is a usable signal for controlling both data augmentation and internal regularization during ASR training. Harder samples get gentler augmentation, so the model can learn from their actual content; easier samples get stronger masking. At the same time, the batch-level average of this complexity modulates the intermediate CTC auxiliary loss, strengthening early-layer regularization when the batch is easy. The two-stage schedule is part of the claim: applying the policy only after an initial converged training phase avoids the large loss fluctuations of early training, which the paper identifies as a cause of extreme and unstable augmentation intensities. The paper reports that this combined scheme outperforms the fixed SpecAugment baseline, a SapAugment re-implementation, and intermediate-CTC regularization alone on both an English and a Mandarin benchmark.

Load-bearing premise

The load-bearing premise is that the min-max normalized training loss of a sample is a stable and calibrated measure of its complexity during the second training stage; if early-stage fluctuations persist or if batch losses cluster tightly, the normalization stops distinguishing samples and the adaptive signals become arbitrary.

Editorial extensions

If this is right

  • If the central claim is right, low-resource ASR systems can be improved by a training-only recipe that leaves the deployed model unchanged in size, speed, or decoding.
  • The two-stage schedule implies that adaptive policies degrade early training; practitioners should warm up with fixed augmentation and regularization before enabling loss-dependent control.
  • The adaptive intermediate-CTC weight provides a new use for intermediate-layer losses: not just a fixed auxiliary loss, but a batch-complexity-dependent regularizer.
  • Ablations in the paper indicate each component, adaptive augmentation and adaptive regularization, contributes independently, and their combination gives the largest gain.

Reading between the lines

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

  • The min-max normalization makes the policy sensitive to batch composition; if batches are homogeneous, nearly every sample gets a policy near the extremes, so the method's gains may depend on the diversity of loss values within each mini-batch.
  • The same batch-complexity scalar could be plugged into other auxiliary losses, such as knowledge-distillation losses or layer-wise losses, beyond intermediate CTC, or into decoder-side regularization.
  • A natural testable extension is to compare MinMax-IBF against gradient-norm or prediction-confidence complexity measures; if loss-based complexity is the right signal, other measures should not improve on it, and if they do, the mechanism is not specifically about loss.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a two-stage training recipe called complexity-boosted adaptive (CBA) training for low-resource ASR. Stage 1 trains a WeNet Conformer with SpecAugment and intermediate-CTC loss regularization. Stage 2 continues training with the proposed MinMax-IBF policy, which normalizes per-sample losses via min-max scaling and uses an incomplete beta function to set an augmentation strength and a batch-wise InterCTC loss weight. On LibriSpeech 100h, the method reports 7.24% and 20.04% WER on test-clean/test-other versus 8.36%/23.34% for the SpecAugment baseline; on AISHELL-1 it reports 4.33% CER versus 4.62%.

Significance. If the reported gains are attributable to the adaptive policy, the work is a useful empirical contribution: a training-only modification with no inference-time overhead, demonstrated on two languages/scripts and packaged as a simple recipe compatible with a widely used toolkit. The paper's strengths are the clear description of the two-stage procedure, the inclusion of SapAugment as a comparative baseline, and ablation experiments on both datasets that show consistent monotonic improvements as components are added. No code or configuration files are released, and all results come from single training runs, so independent verification is not yet possible.

major comments (4)
  1. [Section 4.4, Tables 4 and 6] The paper attributes the 13.4% and 14.1% relative WER improvements to the MinMax-IBF adaptive policy, but the ablation design does not control for the two-stage continued-training schedule. Row 3 of Table 4 ('Regularization(R)') is described as non-adaptive but it is not reported whether it is trained from scratch in a single stage or with a continuation; Table 6 compares from-scratch with continued training only for the adaptive configurations (FS vs CT rows), and never includes a non-adaptive continuation where the stage-1 model is fine-tuned with fixed SpecAugment and fixed InterCTC for the same number of additional epochs. Without this control, the gain of CBA over the Regularization baseline could come from additional fine-tuning, a lower stage-2 learning rate, or continued SpecAugment exposure rather than from the complexity signal. This control is necessary to support the central mechanistic claim.
  2. [Section 4.2 and Section 3.3] Several hyperparameters that are essential for reproducing the recipe and for attributing the effect are omitted: the InterCTC weight lambda in Eqs. (5) and (9), the number of epochs for stage 1 versus stage 2, the learning rate and any schedule change at the beginning of stage 2, and the exact linear mapping from the policy value fDA to the number of time and frequency masks (Section 3.1 gives an example of four time masks but no formula or range). Reporting these values, together with the stage-2 model averaging details, is a prerequisite for verifying the results and for comparing the adaptive recipe against the non-adaptive continuation control requested in the previous comment.
  3. [Section 3.1, Eq. (6)] The proposed MinMax-IBF policy relies on a min-max normalization that may be unstable under exactly the conditions the authors themselves caution about in Section 3.3. When the losses within a mini-batch are nearly equal, xi values collapse toward 0 or 1 and the policy prescribes extreme augmentation and regularization intensities; when a single outlier dominates, all other samples map to near-zero xi and receive nearly maximal augmentation. The paper provides no analysis of the stage-2 loss distributions, no sensitivity experiments on batch size or normalization variants, and no criterion for deciding when the model is 'sufficiently converged' for the policy to be reliable. Because this normalization is the core of the contribution, the stability and calibration of Eq. (6) should be demonstrated empirically.
  4. [Tables 3-6] All experimental numbers come from single training runs with no error bars or significance tests. Several comparisons in the ablation tables are of small magnitude (e.g., 8.30 vs 7.89 on LibriSpeech test-clean in Table 3; 4.62 vs 4.60 on AISHELL-1 in Table 3), so without variance estimates or multiple seeds the reader cannot judge whether the component-wise improvements are reliable. Reporting the mean and standard deviation over at least three seeds for the key rows would materially strengthen the central claim.
minor comments (5)
  1. [Table 3 caption] The caption says 'CER performance' but the table reports WER for LibriSpeech and CER for AISHELL-1; please change the caption to 'WER/CER performance' and correct the misspelling of LibriSpeech.
  2. [Section 3.1] The text following Eq. (6) uses 'min-batch' instead of 'mini-batch'.
  3. [Section 4.2] The batch size used for the MiniMax-IBF policy and for training is not specified; please report it together with the IBF parameters.
  4. [Section 4.2] Please clarify whether layer 12 in the 'subModel selecting layers 6, 9 and 12' is the final encoder output; if so, including it in Eq. (10) duplicates the main CTC loss.
  5. [Section 4.3] The sentence 'achieve WER 7.24% and 20.04%' should read 'achieve WERs of 7.24% and 20.04%'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CBA training is an empirical recipe whose MinMax-IBF policy is an online loss-feedback heuristic, validated on held-out public test sets. The flagged concerns are ablation-schedule confounds and missing hyperparameter details, which are correctness risks, not circular steps.

full rationale

The paper's claimed chain is: (i) define a per-sample complexity signal xi=(Li-Lmin)/(Lmax-Lmin) from the current batch's losses (Eq. 6); (ii) map xi through the incomplete beta function to an augmentation intensity fDA=1-IBF(xi) (Eq. 7) and an InterCTC scalar fCTC=mean(1-IBF(xi)) (Eq. 8); (iii) train non-adaptively in stage 1, then continue with the adaptive policy in stage 2, evaluating WER/CER on the held-out LibriSpeech test-clean/test-other and AISHELL-1 test sets. None of these steps reduces to its inputs. The policy is an online feedback rule: the extra forward pass reads current-batch losses and shapes the next update's augmentation and regularization; the loss is a training signal, not a parameter fitted to the evaluation data. The headline 13.4%/14.1% relative WER gains are measured on public test sets the policy never saw, so they cannot be forced by construction. There are no self-citations: the method builds on SapAugment [6], intermediate CTC [19], and WeNet [21], all external works, and the IBF hyperparameters (s=0.5, a=5) are explicitly inherited from [6], not smuggled in via a same-author citation. No uniqueness theorem is invoked, and no prior result by these authors is load-bearing. The paper itself flags the policy's key limitation in Section 3.3, noting that early-training loss fluctuations would make MinMax-IBF unreliable, hence the non-adaptive first stage; this is an acknowledged constraint, not a circular patch. The genuine concerns are experimental-design confounds, not circularity: Table 4 never runs a 'fixed policy, same two-stage schedule' baseline, so the two-stage fine-tuning schedule's contribution is not isolated from the adaptive policy's, and the stage-1/stage-2 epoch split and stage-2 learning rate are omitted. Table 6 compares from-scratch versus continued training only for the adaptive recipes. These are correctness and reproducibility risks that do not make the central claim circular, because nothing is fitted to the test data and then re-reported as a prediction. Verdict: no significant circularity; score 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. The central claim rests on several domain assumptions inherited from SapAugment and intermediate CTC literature, plus hand-chosen hyperparameters, some unreported. The most fragile assumption is that min-max normalized batch loss is a stable difficulty signal in the second stage.

free parameters (6)
  • IBF shape parameters (s, a) = s=0.5, a=5
    Inherited from SapAugment without retuning; controls the policy curve in Eq. (7) and Eq. (8).
  • InterCTC weight lambda = not reported
    In Eq. (9), lambda balances final CTC loss and intermediate CTC loss. No value is given in the paper, and no sensitivity analysis is provided.
  • Time mask count scaling factor = 4
    Sec 3.1 states 'the number of time masks can be calculated as 4 times the value of fDA'; the factor 4 and the analogous frequency-mask mapping are chosen by hand.
  • Intermediate sub-model layer indices = 6, 9, 12
    Sec 4.2 selects layers 6, 9, 12 from the 12-layer encoder with no ablation or justification for this choice.
  • Stage 1 / stage 2 epoch split = not reported
    The paper says 'continue training for certain epochs' (Sec 3.3) but never reports how the 120 (LibriSpeech) or 240 (AISHELL-1) epochs are divided between the non-adaptive and adaptive stages.
  • Number of averaged models = 10
    Sec 3.3 states 'average 10 models selected by cross-validation loss'; the choice of 10 is arbitrary and unablated.
assumptions (4)
  • standard math The incomplete beta function IBF(s,a; x) maps x monotonically to a policy value in [0,1].
    Used in Eq. (7) and Eq. (8); standard special function, no proof needed.
  • domain assumption A sample's current loss is a valid proxy for its difficulty, and difficulty should inversely control augmentation intensity.
    Borrowed from SapAugment (Sec 2); the entire adaptive policy depends on this premise. The paper does not test alternatives.
  • domain assumption Loss values within a mini-batch are comparable, and min-max normalization yields stable complexity estimates in the second training stage.
    Sec 3.1 Eq. (6); the paper itself notes instability in the first stage (Sec 3.3) but assumes the normalized loss is reliable later.
  • domain assumption Averaging intermediate CTC losses from sub-models at layers 6, 9, 12 regularizes the encoder beneficially.
    Eq. (10) and Sec 4.2; the choice of these layers is not justified and the effect could vary by architecture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Complexity boosted adaptive training for better low resource ASR performance." pith.science (2026). https://pith.science/paper/PSOFMZP3

@misc{pith2026241200877,
  author       = {Pith},
  title        = {Pith review of: Complexity boosted adaptive training for better low resource ASR performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSOFMZP3}},
  note         = {Machine review of arXiv:2412.00877}
}
read the original abstract

During the entire training process of the ASR model, the intensity of data augmentation and the approach of calculating training loss are applied in a regulated manner based on preset parameters. For example, SpecAugment employs a predefined strength of augmentation to mask parts of the time-frequency domain spectrum. Similarly, in CTC-based multi-layer models, the loss is generally determined based on the output of the encoder's final layer during the training process. However, ignoring dynamic characteristics may suboptimally train models. To address the issue, we present a two-stage training method, known as complexity-boosted adaptive (CBA) training. It involves making dynamic adjustments to data augmentation strategies and CTC loss propagation based on the complexity of the training samples. In the first stage, we train the model with intermediate-CTC-based regularization and data augmentation without any adaptive policy. In the second stage, we propose a novel adaptive policy, called MinMax-IBF, which calculates the complexity of samples. We combine the MinMax-IBF policy to data augmentation and intermediate CTC loss regularization to continue training. The proposed CBA training approach shows considerable improvements, up to 13.4% and 14.1% relative reduction in WER on the LibriSpeech 100h test-clean and test-other dataset and also up to 6.3% relative reduction on AISHELL-1 test set, over the Conformer architecture in Wenet.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 25 canonical work pages

  1. [1]

    Besides the model architecture, the performance of ASR systems relies heavily on the availability of large and di- verse training data with transcripts

    INTRODUCTION Automatic Speech Recognition (ASR) has witnessed remarkable ad- vancements in recent years by well-designed self-attention-based End-to-end models. Besides the model architecture, the performance of ASR systems relies heavily on the availability of large and di- verse training data with transcripts. The performance often comes with great comp...

    work page

  2. [2]

    It introduces feature domain time masking and fre- quency masking to data augmentation

    RELA TED WORK SpecAugment. It introduces feature domain time masking and fre- quency masking to data augmentation. It simply masks out segments of sizes mt, and mf , along the time and the frequency axes, respec- tively. SpecAugment is applied to Listen, Attend, and Spell net- works for end-to-end speech recognition tasks, surpassing all previ- ous studie...

    work page

  3. [3]

    The operation of the mask remains at twice in WeNet Conformer

    PROPOSED ADAPTIVE TRAINING METHOD SpecAugment employs a fixed strategy without adaptive adjustments based on the samples. The operation of the mask remains at twice in WeNet Conformer. SapAugment takes initial steps to tackle this is- sue. However, the policy based on loss rank fails to consider the vari- ability of individual samples within a batch, as m...

    work page

  4. [4]

    Dataset We perform our experiments on two datasets: AISHELL-1 [23] and LibriSpeech [24]

    EXPERIMENTS & ANALYSIS 4.1. Dataset We perform our experiments on two datasets: AISHELL-1 [23] and LibriSpeech [24]. AISHELL-1 is an open-source Mandarin speech corpus containing 120,098 utterances from 340 speakers in the train- ing set; the development set contains 14,326 utterances from 40 speakers; the test set contains 7,176 utterances from 20 speake...

    work page

  5. [5]

    Ini- tially, we apply regularization to the training process by utilizing the intermediate CTC loss regularization

    CONCLUSION This paper focuses on a complexity-boosted adaptive training method for better ASR performance in low-resource scenarios. Ini- tially, we apply regularization to the training process by utilizing the intermediate CTC loss regularization. Subsequently, we pro- ceed with training using the proposed MinMax-IBF policy for adap- tive data augmentati...

    work page

  6. [6]

    Sapaugment: Learning a sample adaptive policy for data augmentation,

    Ting-Yao Hu, Ashish Shrivastava, Jen-Hao Rick Chang, Hema Koppula, Stefan Braun, Kyuyeon Hwang, Ozlem Kalinli, and Oncel Tuzel, “Sapaugment: Learning a sample adaptive policy for data augmentation,” in ICASSP 2021-2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 4040–4044

    work page 2021

  7. [7]

    Specaugment: A simple data augmentation method for automatic speech recog- nition,

    Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le, “Specaugment: A simple data augmentation method for automatic speech recog- nition,” arXiv preprint arXiv:1904.08779, 2019

    arXiv 1904

  8. [8]

    Listen, attend and spell: A neural network for large vocab- ulary conversational speech recognition,

    William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals, “Listen, attend and spell: A neural network for large vocab- ulary conversational speech recognition,” in 2016 IEEE inter- national conference on acoustics, speech and signal processing (ICASSP). IEEE, 2016, pp. 4960–4964

    work page 2016

Show all 31 references
  1. [9]

    Attention-based mod- els for speech recognition,

    Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio, “Attention-based mod- els for speech recognition,” Advances in neural information processing systems, vol. 28, 2015

    2015

  2. [10]

    Making more of little data: Improving low-resource automatic speech recognition using data augmen- tation,

    Martijn Bartelds, Nay San, Bradley McDonnell, Dan Jurafsky, and Martijn Wieling, “Making more of little data: Improving low-resource automatic speech recognition using data augmen- tation,” arXiv preprint arXiv:2305.10951, 2023

    2023 arXiv

  3. [11]

    Asr data aug- mentation in low-resource settings using cross-lingual multi- speaker tts and cross-lingual voice conversion,

    Edresson Casanova, Christopher Shulby, Alexander Korolev, Arnaldo Candido Junior, Anderson da Silva Soares, San- dra Alu ´ısio, and Moacir Antonelli Ponti, “Asr data aug- mentation in low-resource settings using cross-lingual multi- speaker tts and cross-lingual voice conversi...

    2022 arXiv

  4. [12]

    On-the-fly aligned data augmentation for sequence- to-sequence asr,

    Tsz Kin Lam, Mayumi Ohta, Shigehiko Schamoni, and Stefan Riezler, “On-the-fly aligned data augmentation for sequence- to-sequence asr,” arXiv preprint arXiv:2104.01393, 2021

    2021 arXiv

  5. [13]

    An inves- tigation of deep neural networks for noise robust speech recog- nition,

    Michael L. Seltzer, Dong Yu, and Yongqiang Wang, “An inves- tigation of deep neural networks for noise robust speech recog- nition,” in 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, 2013, pp. 7398–7402

    2013

  6. [14]

    A study on data augmen- tation of reverberant speech for robust speech recognition,

    Tom Ko, Vijayaditya Peddinti, Daniel Povey, Michael L. Seltzer, and Sanjeev Khudanpur, “A study on data augmen- tation of reverberant speech for robust speech recognition,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 5220–5224

    2017

  7. [15]

    Squeezeformer: An efficient transformer for automatic speech recognition,

    Sehoon Kim, Amir Gholami, Albert Shaw, Nicholas Lee, Kart- tikeya Mangalam, Jitendra Malik, Michael W Mahoney, and Kurt Keutzer, “Squeezeformer: An efficient transformer for automatic speech recognition,” Advances in Neural Informa- tion Processing Systems, vol. 35, pp. 9361–9...

    2022

  8. [16]

    Various aux- iliary loss functions have been explored for different architectures

    and attention-based encoder-decoder (AED) [17]. Various aux- iliary loss functions have been explored for different architectures

  9. [17]

    E-branchformer: Branchformer with enhanced merging for speech recognition,

    Kwangyoun Kim, Felix Wu, Yifan Peng, Jing Pan, Prashant Sridhar, Kyu J Han, and Shinji Watanabe, “E-branchformer: Branchformer with enhanced merging for speech recognition,” in 2022 IEEE Spoken Language Technology Workshop (SLT) . IEEE, 2023, pp. 84–91

    2022

  10. [18]

    However, the model’s training strategy remains rigid re- gardless of the complexity of training samples

    [19]. However, the model’s training strategy remains rigid re- gardless of the complexity of training samples. In this paper, we adopt a two-stage training approach, as shown in Figure 1. Firstly, we propose the MinMax-IBF policy that calcu- lates the complexity of samples, le...

    2024 arXiv

  11. [19]

    Improving mandarin speech recogntion with block- augmented transformer,

    Xiaoming Ren, Huifeng Zhu, Liuwei Wei, Minghui Wu, and Jie Hao, “Improving mandarin speech recogntion with block- augmented transformer,” arXiv preprint arXiv:2207.11697 , 2022

    2022 arXiv

  12. [20]

    Efficient conformer: Progressive downsampling and grouped attention for automatic speech recognition,

    Maxime Burchi and Valentin Vielzeuf, “Efficient conformer: Progressive downsampling and grouped attention for automatic speech recognition,” in 2021 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU) . IEEE, 2021, pp. 8–15

    2021

  13. [21]

    Audio augmentation for speech recognition,

    Tom Ko, Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khu- danpur, “Audio augmentation for speech recognition,” in Six- teenth annual conference of the international speech commu- nication association, 2015

    2015

  14. [22]

    Connectionist temporal classification: la- belling unsegmented sequence data with recurrent neural net- works,

    Alex Graves, Santiago Fern ´andez, Faustino Gomez, and J¨urgen Schmidhuber, “Connectionist temporal classification: la- belling unsegmented sequence data with recurrent neural net- works,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

    2006

  15. [23]

    Sequence transduction with recurrent neural networks,

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

    2012 arXiv

  16. [24]

    Attention is all you need,

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  17. [25]

    Multitask learning with low-level auxiliary tasks for encoder-decoder based speech recognition,

    Shubham Toshniwal, Hao Tang, Liang Lu, and Karen Livescu, “Multitask learning with low-level auxiliary tasks for encoder-decoder based speech recognition,” arXiv preprint arXiv:1704.01631, 2017

    2017 arXiv

  18. [26]

    Intermediate loss regular- ization for ctc-based speech recognition,

    Jaesong Lee and Shinji Watanabe, “Intermediate loss regular- ization for ctc-based speech recognition,” in ICASSP 2021- 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6224–6228

    2021

  19. [27]

    Conformer: Convolution- augmented transformer for speech recognition,

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Par- mar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zheng- dong Zhang, Yonghui Wu, et al., “Conformer: Convolution- augmented transformer for speech recognition,”arXiv preprint arXiv:2005.08100, 2020

    2005 arXiv

  20. [28]

    Wenet: Production oriented streaming and non-streaming end-to-end speech recognition toolkit,

    Zhuoyuan Yao, Di Wu, Xiong Wang, Binbin Zhang, Fan Yu, Chao Yang, Zhendong Peng, Xiaoyu Chen, Lei Xie, and Xin Lei, “Wenet: Production oriented streaming and non-streaming end-to-end speech recognition toolkit,” arXiv preprint arXiv:2102.01547, 2021

    2021 arXiv

  21. [29]

    Paris, Incomplete beta functions , NIST Handbook of Mathematical Functions, Cambridge University Press, 2010

    R.B. Paris, Incomplete beta functions , NIST Handbook of Mathematical Functions, Cambridge University Press, 2010

    2010

  22. [30]

    Aishell-1: An open-source mandarin speech corpus and a speech recognition baseline,

    Hui Bu, Jiayu Du, Xingyu Na, Bengu Wu, and Hao Zheng, “Aishell-1: An open-source mandarin speech corpus and a speech recognition baseline,” in 2017 20th conference of the oriental chapter of the international coordinating committee on speech databases and speech I/O systems an...

    2017

  23. [31]

    Librispeech: an asr corpus based on public do- main audio books,

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “Librispeech: an asr corpus based on public do- main audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2015, pp. 5206–5210

    2015

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.