Pith. sign in

REVIEW 4 major objections 6 minor 33 references

A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions

T0 review · 4 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 Synthetic speech generated from LLM text improves automated scoring of spoken opinion responses, outperforming real-data-only training in low-resource settings.

desk verdict A coherent, honest application of known augmentation tricks to a niche ASA problem, but the headline gain is a 2–3 item difference on tiny test sets with no error bars; worth reviewing, not yet convincing. read the letter →

arxiv 2506.04077 v2 pith:JCB6NBMP submitted 2025-06-04 cs.CL cs.SDeess.AS

Chung-Chun Wang , Jhen-Ke Lin , Hao-Chien Lu , Hong-Yun Lin , Berlin Chen This is my paper
classification cs.CLcs.SDeess.AS
keywords automatedspeakingassessmentopinionexpressiondataaugmentationmultimodallargelanguagemodeltext-to-speechdynamicimportancereweightinglow-resourcesyntheticspeech
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 proposes a training pipeline for automated speaking assessment (ASA) of opinion-expression responses when labeled recordings are scarce. It claims that combining LLM-generated proficiency-conditioned responses, speaker-conditioned TTS that clones real learner voices, and a dynamic importance reweighting loss lets a multimodal large language model score spoken opinion responses more accurately than training on real data alone. On the LTTC GEPT dataset, the full pipeline reaches 76.67% overall accuracy on seen prompts and 64.00% on unseen prompts, compared with 73.33% and 63.33% for the same backbone trained only on real data. If correct, the pipeline offers a way to build low-resource speaking assessment systems without collecting large manually scored corpora.

What carries the argument

The carrying mechanism is a two-stage training loop with a dynamic importance ratio. A frozen quality model $f_Q$ trained on real data and a trainable target model $f_T$ each assign confidence to the correct label, and the per-instance weight is $r_i = p_{Q,i}/p_{T,i}$, multiplied by a real-instance bonus $\alpha > 1$; highly realistic synthetic instances that the quality model scores confidently are amplified, while synthetic outliers are suppressed. This reweighting is attached to a pipeline that generates response text with an LLM, converts it to learner-like audio with a multi-speaker TTS conditioned on a speaker embedding $v_{l,j}$ extracted from a real learner clip, and feeds text plus audio into Phi-4 multimodal for score prediction.

What would settle it

Force the TTS to synthesize all generated responses with one native-speaker voice, keeping everything else identical; if seen and unseen accuracy remain at 76.67% and 64.00%, then speaker-conditioned cloning is not the active ingredient, and the claim that learner-like synthetic audio is necessary would be unsupported.

Watch

Extended reading notes

Core claim

The central claim is that synthetic data generated to mimic real non-native speakers can substitute for a large share of manually scored recordings in automatic scoring of opinion expressions, provided the generation is proficiency-aware and the training loss down-weights less realistic samples. The authors show that training a Phi-4 multimodal backbone with LoRA on a mixture of real and synthesized instances, with real instances weighted twice as heavily, yields the best accuracy in their comparison, and that training on synthesized speech alone collapses to 42.22% seen accuracy. They interpret the gap as evidence that the dynamic importance loss keeps the model grounded in authentic speech while letting diverse synthetic text and cloned voices expand prompt coverage.

Load-bearing premise

The entire gain depends on the multi-speaker TTS cloning real learner voices closely enough that synthesized audio carries the disfluencies, mispronunciations, and prosody of non-native speech; the authors themselves note that accent and hesitation mismatches persist.

Editorial extensions

If this is right

  • ASA systems for opinion-expression tasks could be built from a few hundred labeled recordings plus a much larger machine-generated corpus.
  • The same dynamic reweighting principle could make synthetic data useful in other low-resource speech tasks where a quality model trained on real data can rank synthetic samples.
  • The reported gains on seen prompts suggest that generative augmentation helps most when the test prompts resemble those seen in training, while unseen-prompt transfer is more modest.
  • Because training on synthesized data alone performs badly, the pipeline should be seen as an augmentation of real data rather than a replacement for it.

Reading between the lines

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

  • The same generate-then-reweight recipe could apply to other low-resource multimodal tasks, such as pronunciation feedback or spoken-language understanding, whenever a quality model trained on real data can rank synthetic samples.
  • Because the unseen-prompt gain over the wav2vec2+BERT baseline is a tie (64.00% vs 64.00%) and the absolute gain over real-data-only Phi-4 is 0.67 percentage points, the transferability advantage is the least certain part of the claim; a larger unseen-prompt evaluation would clarify it.
  • The speaker-conditioned cloning step may matter most for seen-prompt accuracy, where the model can memorize prompt-specific content; testing on more prompt types would show whether the diversity from LLM text is doing the heavy lifting.
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

4 major / 6 minor

Summary. The paper proposes a three-stage data augmentation pipeline for automatic speaking assessment of opinion expressions under low-resource conditions: an LLM generates proficiency-conditioned text responses, a speaker-conditioned multi-speaker TTS converts them into learner-like speech, and a dynamic importance reweighting loss adaptively balances synthetic and real instances. The scoring backbone is Phi-4 multimodal fine-tuned with LoRA. Experiments on the LTTC GEPT dataset compare the proposed model with single-modality and dual-modality baselines on a 90-item seen test set and a 300-item unseen prompt test set. The central claim is that the full pipeline outperforms training on real data alone, with reported seen overall accuracy of 76.67% versus 73.33% and unseen overall accuracy of 64.00% versus 63.33%, plus an ablation showing that synthetic-only training performs poorly and that mixing with an elevated real-instance weight is best.

Significance. If the reported gains are real, the paper describes a useful, replicable recipe for low-resource spoken assessment and a sensible integration of LLM text generation, speaker-conditioned TTS, and loss reweighting. The pipeline is internally coherent, the implementation details are reasonably concrete, and the ablation includes a valuable negative control: synthetic-only training collapses to 42.22% seen and 46.67% unseen accuracy, showing that the synthetic data cannot simply replace real speech. The paper also gives specific model choices (o4-mini, XTTSv2, Phi-4 multimodal, LoRA), which aids reproducibility. However, the headline outperformance over real-data-only training rests on very small classification differences—3 of 90 seen items and 2 of 300 unseen items—and the paper reports no confidence intervals, repeated seeds, or significance tests. The stated Limitations also concede that accent and hesitation mismatches persist, which directly tempers the TTS-fidelity premise in §2.2. The significance of the contribution is therefore currently unproven, although the framework itself is worth further evaluation.

major comments (4)
  1. [§3.3, Table 1] The central claim that the proposed approach outperforms methods relying on real data is not statistically supported by the reported numbers. On the 90-item seen test set, 76.67% versus 73.33% corresponds to 69 versus 66 correct classifications, a difference of 3 items. On the 300-item unseen test set, 64.00% versus 63.33% corresponds to 192 versus 190 correct classifications, a difference of 2 items, and the wav2vec2+BERT baseline ties the proposed model at 64.00% on the unseen set. No confidence intervals, multiple seeds, or significance tests are reported anywhere in §3.3. These deltas are within the range of typical run-to-run and optimizer noise for a 720-example training set, so the headline outperformance claim is not yet established.
  2. [§2.3, Eq. (7); §3.2] The training objective is not fully specified and contains a notation collision that prevents exact reproduction. Equation (7) introduces η inside the loss term and calls it the learning rate, which is nonstandard and makes the loss magnitude depend on an optimizer hyperparameter. Section 3.2 then reports 'learning rate α = 2' and separately states 'applying a real-sample weight α = 2', so it is unclear whether α is the learning rate, the real-instance weight, or whether η and α have been conflated. The authors should define a single set of symbols and state whether the learning rate appears only in the optimizer or also as a multiplicative factor in the loss.
  3. [§3.4, Table 2] The ablation's key comparison, Mix (α=1) versus Mix (α=2), does not provide strong evidence for the benefit of elevated real-instance weighting. The seen-set difference is 4.45 percentage points (72.22% versus 76.67%, 65 versus 69 of 90 items), but the unseen-set difference is only 0.33 percentage points (63.67% versus 64.00%, 191 versus 192 of 300 items). With no significance testing or repeated runs, the claim that increasing the real-instance weight yields the best results is supported only by the seen-set gap; the unseen-set difference is a single item and is indistinguishable from noise.
  4. [§2.2; Limitations] The method's rationale depends on speaker-conditioned TTS preserving disfluencies, mispronunciations, and prosodic variability of non-native speech, but the Limitations section concedes that 'accent and hesitation mismatches persist.' The paper provides no acoustic or perceptual evaluation of how close the synthesized learner speech is to real learner speech. Given that the only clearly robust empirical finding is that synthetic-only training performs far worse than mixed training (Table 2), the specific contribution of the TTS-fidelity assumption to the final accuracy is not directly measured or isolated.
minor comments (6)
  1. [§2.2] The phrase 'By randomly pairing each g_{l,i} with different v_{l,j}' uses the undefined symbol g_{l,i}; it should presumably refer to the generated text \hat{R}_{l,i} from Eq. (4).
  2. [§3.2] The synthetic corpus is said to match the size and score distribution of the real training set, but the number n of generated instances per proficiency level and the total synthetic corpus size are not reported; these values should be stated for reproducibility.
  3. [§3.3] The unseen test set consists of 300 triples from the fourth prompt type. Because the paper's contribution is framed around prompt diversity and generalization, it should clarify how many distinct prompts are contained in this unseen set and how the 8:1:1 split of the remaining 900 triples was stratified across prompt types.
  4. [§3.3] There are several typos in the results prose, including 'On the the test', 'surpasseing', and 'alleviate the downsides the systems'; these should be corrected.
  5. [Figure 3] Figure 3 shows only the real-data score distribution. Adding the corresponding distribution for the synthesized corpus would help support the claim that the augmentation matches the real label distribution.
  6. [§2.3] The importance ratio r_i = p_{Q,i}/p_{T,i} may become very large when p_{T,i} is near zero early in training; the paper should state whether r_i is clipped or bounded and how the quality model's confidence scores are calibrated.

Circularity Check

0 steps flagged · score 2.0 of 10

No substantive circularity: the central claim is an empirical comparison against independent held-out real data, and the only minor issues are two non-load-bearing self-citations plus a training-time reweighting loop.

full rationale

The paper's central claim ('our approach outperforms methods relying on real data or conventional augmentation') is supported by Table 1, which compares the full pipeline against Phi-4 trained on real-only data, wav2vec2, BERT, and wav2vec2+BERT on held-out seen and unseen test sets. No fitted parameter is renamed as a prediction: the real-instance weight alpha=2 is fixed in Eq. (7) and not tuned on the test sets, and the final scores come from the target model f_T, which is benchmarked on real, held-out triples. The dynamic importance loss (Eqs. 5-8) does create a mild self-referential loop in that the frozen quality model f_Q, fitted on real data, supplies confidence weights for synthetic instances during f_T's training; however, this is a training procedure, not a derivation that makes the test accuracy equal to an input. f_T still learns from real labels and its accuracy is not forced by any equation. The self-citations ([5] and [7], both sharing author Berlin Chen) appear only as illustrative examples of wav2vec-2.0-based ASA work in the introduction; they are not the source of the augmentation scheme, the dynamic loss, or the scoring backbone, and no uniqueness claim or ansatz is imported from them. The limitations statement about accent/hesitation mismatches and the small Table 1 deltas (3 seen and 2 unseen correct items) are statistical and external-validity concerns, not circularity. Thus no step reduces to its input by construction; the score reflects only the minor self-citations.

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

The approach rests on four hand-chosen hyperparameters and three domain-behavior assumptions about the realism of the synthetic data. It introduces no new theoretical entities. The dynamic importance loss (Eqs. 5-7) is a direct adaptation of Kuo et al. [24], with the alpha weight selected on the test set (Table 2).

free parameters (4)
  • real-instance weight alpha = 2
    Selects the weight of real instances in the dynamic importance loss (Eq. 7); the improvement from alpha=1 to alpha=2 is exactly the headline seen-test gain, so the result depends on this hand-chosen value.
  • LLM decoding temperature tau = 1.5
    Chosen to increase diversity in LLM text generation (Sec. 3.2); no sensitivity analysis is reported.
  • number of in-context examples k = up to 10
    Randomly sampled and permuted examples per prompt (Sec. 2.1); no ablation on k is reported.
  • number of generated instances n = not reported; synthetic corpus matched to real training size
    Synthetic set is described as matching the size and score distribution of the real training set, but n per level is not stated in the paper.
assumptions (4)
  • domain assumption LLM-generated texts at a given proficiency level align with real scoring rubrics.
    Sec. 2.1: generation is conditioned on shuffled real transcripts of that level (Eq. 1), but no human or automatic verification of the generated text labels is reported.
  • domain assumption Speaker-conditioned TTS preserves learner-like disfluencies and prosody.
    Sec. 2.2 claims this, but the Limitations section admits 'accent and hesitation mismatches persist', weakening the premise.
  • domain assumption The quality model f_Q trained on 720 real triplets gives reliable confidence estimates p_Q on synthetic instances.
    Sec. 2.3 Stage 1: f_Q is the reference for reweighting; its reliability on out-of-distribution synthetic data is assumed, not demonstrated.
  • domain assumption Manual transcripts are a faithful proxy for ASR output in deployment.
    Sec. 3.1: all spoken responses were manually transcribed, so ASR errors, which would appear in practice, are not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions." pith.science (2026). https://pith.science/paper/JCB6NBMP

@misc{pith2026250604077,
  author       = {Pith},
  title        = {Pith review of: A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCB6NBMP}},
  note         = {Machine review of arXiv:2506.04077}
}
read the original abstract

Automated speaking assessment (ASA) on opinion expressions is often hampered by the scarcity of labeled recordings, which restricts prompt diversity and undermines scoring reliability. To address this challenge, we propose a novel training paradigm that leverages a large language models (LLM) to generate diverse responses of a given proficiency level, converts responses into synthesized speech via speaker-aware text-to-speech synthesis, and employs a dynamic importance loss to adaptively reweight training instances based on feature distribution differences between synthesized and real speech. Subsequently, a multimodal large language model integrates aligned textual features with speech signals to predict proficiency scores directly. Experiments conducted on the LTTC dataset show that our approach outperforms methods relying on real data or conventional augmentation, effectively mitigating low-resource constraints and enabling ASA on opinion expressions with cross-modal information.

Figures

Figures reproduced from arXiv: 2506.04077 by the authors.

Figure 1
Figure 1. Pipeline for synthesized speech generation. We build in-context prompts from k real transcripts and condition a multi-speaker TTS model on LLM-generated text Rˆl,i and a speaker embedding vl to produce learner-like audio. we introduce a novel dynamic importance loss that builds on a recently proposed data-weighting principle [24] for training of our scoring model. This training loss automatically increases the contr… view at source ↗
Figure 2
Figure 2. Two-stage training workflow. Stage 1 fine-tunes a Quality Model fQ on real data. Stage 2 trains the Target Model fT on mixed real and synthesized instances using the ratio ri and weight wi to dynamically reweight each instance’s loss. traits, in contrast to the uniform, overly clean output generated by a conventional single-speaker TTS model. 2.3. End-to-End Multimodal Training with Dynamic Im￾portance Reweighting A… view at source ↗
Figure 3
Figure 3. Score distribution for real data. We optimize the average of these losses over all N training in￾stances: L = 1 N XN i=1 ℓi. (8) By introducing this novel training loss, we ensure that the augmentation of synthesized instances enriches the training set without distorting the model’s prediction capability on the au￾thentic speech. By doing so, the estimated multimodal scoring model is anticipated to achieve robust pe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    Introduction In recent years, technologies for computer-assisted language learning (CALL), such as automated speaking assessment (ASA), have made significant strides to meet the growing de- mand for scalable and objective evaluation of second-language (L2) speaking proficiency in both academic and professional contexts [1, 2, 3]. In L2 speaking examinatio...

    work page Pith review arXiv 2025

  2. [2]

    Please generate your response in the style of the above examples

    Methodology Our approach enhances opinion expression assessment in low resource settings through three tightly integrated innovations. First, a proficiency aware text synthesis module uses a large language model, prompted with randomly sampled and per- muted transcripts of authentic test taker responses, to gener- ate answers aligned with each proficiency...

    work page

  3. [3]

    Experiments and Results 3.1. Dataset We used the LTTC GEPT Intermediate to Advanced opinion ex- pression dataset, which contains 1 200 prompt–response–score triples evenly distributed across four prompt types within the discussion subtask. Of these, 300 triples belonging to the fourth prompt type were set aside as an unseen test set in order to simulate d...

    work page

  4. [4]

    Conclusion In this paper, we have proposed a unified, low-resource train- ing framework for opinion-expression assessment that inte- grates LLM-driven text synthesis, speaker-conditioned voice cloning, and a novel dynamic importance reweighting loss within an end-to-end multimodal model. Empirical results on the LTTC GEPT Intermediate–Advanced dataset hav...

    work page

  5. [5]

    Any findings and implications in the paper do not necessarily reflect those of the sponsor

    Acknowledgement This work was supported by the Language Training and Testing Center (LTTC), Taiwan. Any findings and implications in the paper do not necessarily reflect those of the sponsor

  6. [6]

    An overview of spoken language technology for education,

    M. Eskenazi, “An overview of spoken language technology for education,”Speech communication, vol. 51, no. 10, pp. 832–844, 2009

    work page 2009

  7. [7]

    Attention-based multi-encoder automatic pronunciation assessment,

    B. Lin and L. Wang, “Attention-based multi-encoder automatic pronunciation assessment,” inICASSP 2021-2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 7743–7747

    work page 2021

  8. [8]

    Towards automatic scoring of a test of spoken language with heterogeneous task types,

    K. Zechner and X. Xi, “Towards automatic scoring of a test of spoken language with heterogeneous task types,” inProceedings of the third workshop on innovative use of NLP for building edu- cational applications, 2008, pp. 98–106

    work page 2008

Show all 33 references
  1. [9]

    Assessment of l2 oral proficiency using self-supervised speech representation learning,

    S. Bann `o, K. M. Knill, M. Matassoni, V . Raina, and M. Gales, “Assessment of l2 oral proficiency using self-supervised speech representation learning,” in9th Workshop on Speech and Lan- guage Technology in Education (SLaTE), 2023, pp. 126–130

    2023

  2. [10]

    An effective automated speaking assessment approach to mitigating data scarcity and imbalanced distribution,

    T.-H. Lo, F.-A. Chao, T.-i. Wu, Y .-T. Sung, and B. Chen, “An effective automated speaking assessment approach to mitigating data scarcity and imbalanced distribution,” inFindings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard, E...

    2024

  3. [11]

    Neural approaches to automated speech scoring of monologue and dialogue responses,

    Y . Qian, P. Lange, K. Evanini, R. Pugh, R. Ubale, M. Mulholland, and X. Wang, “Neural approaches to automated speech scoring of monologue and dialogue responses,” inICASSP 2019-2019 IEEE international conference on acoustics, speech and signal process- ing (ICASSP). IEEE, 201...

    2019

  4. [12]

    Development of an english oral assessment system with the gept dataset,

    H.-C. Lu, C.-C. Wang, J.-K. Lin, and B. Chen, “Development of an english oral assessment system with the gept dataset,” in 2024 27th Conference of the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of Speech Databases and Assessment Techniqu...

    2024

  5. [13]

    Assessment of non-native speech using vowel space characteristics,

    L. Chen, K. Evanini, and X. Sun, “Assessment of non-native speech using vowel space characteristics,” in2010 IEEE Spoken Language Technology Workshop. IEEE, 2010, pp. 139–144

    2010

  6. [14]

    Deep learning-based automatic pronunciation assessment for second language learners,

    K. Takai, P. Heracleous, K. Yasuda, and A. Yoneyama, “Deep learning-based automatic pronunciation assessment for second language learners,” inHCI International 2020-Posters: 22nd In- ternational Conference, HCII 2020, Copenhagen, Denmark, July 19–24, 2020, Proceedings, Part II...

    2020

  7. [15]

    Automatic assessment of second lan- guage learners’ fluency,

    H. Strik and C. Cucchiarini, “Automatic assessment of second lan- guage learners’ fluency,” 1999

    1999

  8. [16]

    Assessing the prosody of non-native speakers of english: Measures and feature sets,

    E. Coutinho, F. H ¨onig, Y . Zhang, S. Hantke, A. Batliner, E. N¨oth, and B. Schuller, “Assessing the prosody of non-native speakers of english: Measures and feature sets,” 2016

    2016

  9. [17]

    A deep learning approach to automatic characterisation of rhythm in non-native english speech

    K. Kyriakopoulos, K. M. Knill, and M. J. Gales, “A deep learning approach to automatic characterisation of rhythm in non-native english speech.” ISCA, 2019

    2019

  10. [18]

    Using bidirectional lstm recurrent neural networks to learn high-level abstractions of sequential features for automated scoring of non- native spontaneous speech,

    Z. Yu, V . Ramanarayanan, D. Suendermann-Oeft, X. Wang, K. Zechner, L. Chen, J. Tao, A. Ivanou, and Y . Qian, “Using bidirectional lstm recurrent neural networks to learn high-level abstractions of sequential features for automated scoring of non- native spontaneous speech,” i...

    2015

  11. [19]

    End-to-end neural network based automated speech scoring,

    L. Chen, J. Tao, S. Ghaffarzadegan, and Y . Qian, “End-to-end neural network based automated speech scoring,” in2018 IEEE international conference on acoustics, speech and signal process- ing (ICASSP). IEEE, 2018, pp. 6234–6238

    2018

  12. [20]

    Automated scoring of nonnative speech using the speechrater sm v. 5.0 engine,

    L. Chen, K. Zechner, S.-Y . Yoon, K. Evanini, X. Wang, A. Louk- ina, J. Tao, L. Davis, C. M. Lee, M. Maet al., “Automated scoring of nonnative speech using the speechrater sm v. 5.0 engine,”ETS Research Report Series, vol. 2018, no. 1, pp. 1–31, 2018

    2018

  13. [21]

    Addressing cold start problem for end-to-end automatic speech scoring,

    J. Park and S. Choi, “Addressing cold start problem for end-to-end automatic speech scoring,” inInterspeech 2023, 2023, pp. 994– 998

    2023

  14. [22]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  15. [23]

    Bert: Pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre- training of deep bidirectional transformers for language under- standing,” inProceedings of the 2019 conference of the North American chapter of the association for computational linguis- tics: human language techno...

    2019

  16. [24]

    Phi-4-mini technical report: Compact yet powerful multi- modal language models via mixture-of-loras,

    A. Abouelenin, A. Ashfaq, A. Atkinson, H. Awadalla, N. Bach, J. Bao, A. Benhaim, M. Cai, V . Chaudhary, C. Chenet al., “Phi-4-mini technical report: Compact yet powerful multi- modal language models via mixture-of-loras,”arXiv preprint arXiv:2503.01743, 2025

    2025 arXiv

  17. [25]

    A knowledge-driven vowel-based ap- proach of depression classification from speech using data aug- mentation,

    K. Feng and T. Chaspari, “A knowledge-driven vowel-based ap- proach of depression classification from speech using data aug- mentation,” inICASSP 2023 - 2023 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  18. [26]

    Text is all you need: Personalizing asr models using controllable speech synthesis,

    K. Yang, T.-Y . Hu, J.-H. R. Chang, H. S. Koppula, and O. Tuzel, “Text is all you need: Personalizing asr models using controllable speech synthesis,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  19. [27]

    Synthasr: Unlocking synthetic data for speech recognition,

    A. Fazel, W. Yang, Y . Liu, R. Barra-Chicote, Y . Meng, R. Maas, and J. Droppo, “Synthasr: Unlocking synthetic data for speech recognition,” inInterspeech 2021, 2021, pp. 896–900

    2021

  20. [28]

    Emomix: Emotion mixing via diffusion models for emotional speech syn- thesis,

    H. Tang, X. Zhang, J. Wang, N. Cheng, and J. Xiao, “Emomix: Emotion mixing via diffusion models for emotional speech syn- thesis,” inInterspeech 2023, 2023, pp. 12–16

    2023

  21. [29]

    Not all LLM-generated data are equal: Rethinking data weighting in text classification,

    H.-Y . Kuo, Y .-H. Liao, Y .-C. Chao, W.-Y . Ma, and P.-J. Cheng, “Not all LLM-generated data are equal: Rethinking data weighting in text classification,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/...

    2025

  22. [30]

    GPT3Mix: Leveraging large-scale language models for text augmentation,

    K. M. Yoo, D. Park, J. Kang, S.-W. Lee, and W. Park, “GPT3Mix: Leveraging large-scale language models for text augmentation,” inFindings of the Association for Computational Linguistics: EMNLP 2021, M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, Eds. Punta Cana, Dominican...

    2021

  23. [31]

    Data augmentation for intent classification with off-the-shelf large language models,

    G. Sahu, P. Rodriguez, I. Laradji, P. Atighehchian, D. Vazquez, and D. Bahdanau, “Data augmentation for intent classification with off-the-shelf large language models,” inProceedings of the 4th Workshop on NLP for Conversational AI, B. Liu, A. Papangelis, S. Ultes, A. Rastogi,...

    2022

  24. [32]

    Self-instruct: Aligning language models with self-generated instructions,

    Y . Wang, Y . Kordi, S. Mishra, A. Liu, N. A. Smith, D. Khashabi, and H. Hajishirzi, “Self-instruct: Aligning language models with self-generated instructions,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), A...

    2023

  25. [33]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”ICLR, vol. 1, no. 2, p. 3, 2022

    2022

Pith tools

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