Pith. sign in

REVIEW 5 major objections 5 minor 34 references

Language-Agnostic Suicidal Risk Detection Using Large Language Models

T0 review · 5 major / 5 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 This paper claims that suicidal-risk features extracted by LLMs from Chinese speech keep their predictive power when rewritten in English, enabling language-agnostic risk detection.

desk verdict The bilingual extraction pipeline is new and the dev numbers are honest, but the cross-lingual claim is undermined by an architecture confound that a matched-pair experiment could fix. read the letter →

arxiv 2505.20109 v1 pith:6RVQTBIN submitted 2025-05-26 cs.CL cs.AI

June-Woo Kim , Wonkyo Oh , Haram Yoon , Sung-Hoon Yoon , Dae-Jin Kim , Dong-Ho Lee , Sang-Yeol Lee , Chan-Mo Yang This is my paper
classification cs.CLcs.AI
keywords suicidalriskdetectionlanguage-agnosticlargelanguagemodelscross-lingualfeatureextractionadolescentmentalhealthautomaticspeechrecognitionmultimodalfusionprompt-basedsummarization
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

This paper proposes a way to detect suicidal risk in adolescent speech without building a separate model for each language. The pipeline transcribes Mandarin speech with an automatic speech recognition system, asks a large language model to summarize the parts of the transcript related to suicidal thoughts, and keeps the summary in both Chinese and English. Each summary is used to fine-tune a pretrained language model in its own language. On the SpeechWellness dev set, models trained on the English summaries perform as well as or better than models trained on Chinese summaries, most clearly on the emotion-description task. The authors take this as evidence that the suicidal-risk signal is carried by the content of what is said, not by the language it is said in.

What carries the argument

The central mechanism is bilingual suicidal risk-related feature extraction: a carefully worded prompt asks a large language model to summarize, in the first person, the key points related to the person's suicidal thoughts from the ASR transcript, producing separate Chinese and English versions. These summaries are then used to fine-tune language-specific pretrained models independently. The cross-lingual transfer is attributed to the summaries retaining the risk-relevant content while discarding the original language's surface form, so that English models can learn from Chinese speech data.

What would settle it

If a language model trained on English summaries produced by a neutral, non-suicidal summarization prompt performed equally well, the transfer would not be attributable to risk-feature extraction. More directly, applying the same pipeline to a second source language and finding that English features no longer match local-language features would refute the language-agnostic claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that suicidal risk indicators extracted from Chinese speech transcripts by GPT-4o or Qwen-Plus remain effective after being rewritten in English, and that fine-tuning English-language models on these English summaries yields performance comparable to fine-tuning Chinese models on the original Chinese features. This is presented in Section 4.1 as the key finding: models trained on English features outperform those trained on Chinese features on the emotion-description (ED) task, demonstrating cross-linguistic transfer without performance degradation. The paper contrasts this with direct fine-tuning on raw ASR transcripts, which scores lower, and with baseline multimodal systems, which the proposed text-plus-speech fusion also exceeds on the dev set. The authors acknowledge a dev-to-test performance drop for their top models, attributing it to possible distribution shift or overfitting.

Load-bearing premise

The load-bearing assumption is that the English summaries preserve the same suicidal-risk cues as the original Chinese transcripts, without dropping or adding information during translation or through the wording of the prompt.

Editorial extensions

If this is right

  • Fine-tuning on LLM-extracted risk features outperforms fine-tuning on raw ASR transcripts for both Chinese and English models, e.g., GPT-CBERT reaches 73% accuracy on ER versus 57% for CBERT.
  • English features extracted from Chinese speech transfer without degradation, matching or beating Chinese-feature-trained models on the ED task, so the framework does not require a per-language annotated dataset.
  • Multimodal fusion of speech representations (via XLSR) with text features beats unimodal baselines, with Qwen-CBERT+XLSR reaching the best combined dev accuracy of 68% and F1 of 69.23%.
  • The choice of extraction LLM matters: GPT-4o features work better for emotional regulation while Qwen-Plus features work better for emotion description, suggesting complementary signal across models.
  • The sharp dev-to-test drop (68% to 54% for the best dev model) indicates that the gains are not yet stable across data distributions and would need further validation.

Reading between the lines

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

  • If the cross-lingual transfer is real, English could serve as a pivot language: any language with an ASR system could be funneled through English summaries to train risk detectors, bypassing the need for new annotated mental-health corpora in each target language.
  • A direct test of the claim would be to run the same pipeline on a second source language, such as Korean or Spanish, and check whether English summaries still match local-language features; a failure there would bound the generalizability.
  • The prompt explicitly directs attention to suicidal thoughts, so an important control is whether generic summarization without that instruction yields the same transfer, which would help rule out the model simply echoing prompt content rather than extracting real risk cues.
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

5 major / 5 minor

Summary. The paper proposes a language-agnostic pipeline for suicidal risk detection in Chinese adolescent speech. Raw speech is transcribed with Paraformer, and GPT-4o or Qwen-Plus are prompted to extract suicidal-risk-related features in both Chinese and English. These features are used to fine-tune language-specific pretrained models (CBERT and UER-Dianping for Chinese; DepRoBERTa, Suicidality-ELECTRA, and Suicide-RoBERTa for English), and the text representations are fused with speech representations from XLSR or other speech encoders. Experiments on the 1st SpeechWellness Challenge dataset are reported on the dev set for text-only, speech-only, and multimodal systems, with test results for two submitted systems. The central claim is that models trained on English features extracted from Chinese transcripts outperform those trained on Chinese features in the Expression Description task, suggesting that suicidal-risk indicators can transfer across languages.

Significance. The problem addressed—language-portable suicidal risk detection in adolescents—is practically important, and the paper provides a broad empirical sweep: two LLM extractors, five language models, three speech models, and multimodal fusion. The authors also include a limitation statement about the MINI-KID label definition, which is commendable. If the cross-lingual claim were rigorously established, the framework would be valuable because it would reduce the need for language-specific resources. However, the experimental design does not currently isolate the effect of feature language from the choice of pretrained model architecture, and the test-set results are markedly worse than the dev-set results. Thus the contribution is promising but not yet convincingly demonstrated.

major comments (5)
  1. [Section 4.1, Table 3] The claim that models trained on English features outperform those trained on Chinese features in the ED task is confounded by model architecture and task specialization. The English models (DepRoBERTa, Suicidality-ELECTRA, Suicide-RoBERTa) are fine-tuned on depressive or suicide-related text, whereas the Chinese models (CBERT, UER-Dianping-RoBERTa) are generic or trained on restaurant reviews. For example, GPT-DepRoBERTa achieves 64% Acc in ED versus 59% for Qwen-CBERT, but this difference could stem from the English models being better adapted to suicidal language rather than from the features being in English. To support the 'language-agnostic' claim, the comparison must hold the model architecture and pretraining/fine-tuning history constant across feature languages (e.g., a multilingual backbone or identical model families).
  2. [Section 4.4, Table 6] The test-set results do not support the paper's claim of comparable performance to baselines. Qwen-CBERT+XLSR drops from 68% dev accuracy to 54% test accuracy, and GPT-DepRoBERTa+XLSR drops from 61% to 47%, while the W2V2+BERT baseline achieves 61% test accuracy. The manuscript reports test results for only two of the proposed systems and nonetheless concludes that the method is 'comparable to or better than' baselines. The abstract and conclusion should be tempered, and test-set results for all systems, or at least a clear explanation of why they are omitted, should be provided.
  3. [Section 4.1, Table 3] No confidence intervals, significance tests, or multiple-run variabilities are reported. Many of the differences used to support the 'outperform' claim are small (e.g., 59% vs 60% accuracy; 64.52 vs 65.55 F1). With a dev set of roughly 100 samples (a 1/6 split of 600), these differences are likely within chance. The authors should include error bars or a paired significance test (e.g., McNemar's test for predictions) before asserting that one method outperforms another.
  4. [Section 4.1 vs Section 3.1] The manuscript describes a 'DepRoBERTa trained on direct English translations of ASR results' condition in the results, but the method section never says that direct English translations were generated or how they were produced. This missing baseline condition is important because it serves as a lower-bound comparison for the LLM-extracted English features. Without a description of the translation process (e.g., which MT system or LLM was used), the experiments are not reproducible and the contribution of the LLM feature extraction step is unclear.
  5. [Section 3.1.2] The prompt explicitly instructs the LLM to 'summarize the key points that might be related to the person's suicidal thoughts' and to write from a first-person perspective. This instruction will naturally bias the extracted summaries toward suicidal cues and away from neutral content, making the features highly curated rather than faithful summaries. The conclusion acknowledges that more general prompts should be explored in future work, but the current results should be interpreted with this limitation in mind. The 'language-agnostic' story would be much stronger if the same pipeline were evaluated with a neutral summarization prompt.
minor comments (5)
  1. [Eq. (1)] The definition of F1 does not specify which class is treated as positive (at-risk or no-risk). Since the choice affects the reported F1 values, please clarify the positive class.
  2. [Table 2] The heading 'GIVEN ASR EXAMPLE OF ER (MISINFORMATION)' is confusing; the term 'MISINFORMATION' is not defined or used elsewhere in the paper.
  3. [Section 3.2 and Table 4] The model is referred to as 'wav2vec2-Large' in the text but as 'wav2vec2-Chinese' in Table 4; please use consistent naming throughout.
  4. [Section 2.1] The statement that the work 'mainly focus[es] on the dev sets' is unusual for a predictive modeling paper; please justify this choice or include test results for all compared methods.
  5. [General] The title and abstract use 'language-agnostic' although only Chinese-to-English transfer is evaluated. Please either broaden the evaluation to include other language pairs or soften the claim to 'cross-lingual (Chinese-English) transfer.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons, not derivations that reduce to their inputs by construction.

full rationale

The paper's load-bearing results are empirical comparisons (Tables 3-6) of dev/test performance after fine-tuning on LLM-extracted summaries, not derivations. The LLM prompts explicitly ask for 'key points that might be related to the person's suicidal thoughts,' but these summaries are generated from ASR transcripts without access to MINI-KID labels, and the risk classifier is then trained and evaluated on the actual labels; thus the feature extraction is a form of prompted feature engineering rather than a definitional identification of features with labels. No equation or construction in the paper equates extracted features with the target label. The cross-lingual comparison in Section 4.1 confounds feature language with model architecture (English classifiers such as DepRoBERTa are task-specialized while Chinese classifiers are generic), but this is a validity/confounding concern, not circularity. The only self-citation ([4], a general background review of suicide risk tools) is not load-bearing. The paper also reports test-set results for the leaderboard, providing external validation against the challenge's held-out set, which further supports the empirical nature of the claims.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new theoretical entity is introduced; the ledger lists only the domain assumptions the empirical claim rests on. Hyperparameters are conventional training settings and are listed for completeness, not because they are fitted to the target result.

free parameters (1)
  • Training hyperparameters = 10 epochs; lr 1e-5 (speech), 5e-5 (text), 1e-3 (fusion); batch sizes 8, 16, 32
    Hand-chosen in Section 2.2; standard settings that affect all reported scores but are not fitted to the target result.
assumptions (5)
  • domain assumption The SW1 dev set is representative of the target adolescent population.
    All comparisons in Section 4 use the dev set; if it is biased, the comparable performance claim does not generalize.
  • domain assumption The MINI-KID labels are accurate enough as ground truth for current suicide risk.
    Section 5.1 limits the labels to self-reported, fixed-item assessment; the model predictions are evaluated against those labels.
  • domain assumption Paraformer ASR preserves the linguistic content relevant to suicidal risk.
    Section 3.1.1 chooses Paraformer based on downstream accuracy; noisy transcription would corrupt the LLM summaries.
  • domain assumption GPT-4o and Qwen-Plus English summaries preserve risk-relevant semantics of the Chinese transcripts.
    The cross-lingual claim in Section 4.1 requires this; no human evaluation or translation quality measurement is reported.
  • domain assumption No label information leaks into the LLM summaries through the prompts.
    The prompts ask for key points that might be related to suicidal thoughts without revealing the label; leakage would inflate reported performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language-Agnostic Suicidal Risk Detection Using Large Language Models." pith.science (2026). https://pith.science/paper/6RVQTBIN

@misc{pith2026250520109,
  author       = {Pith},
  title        = {Pith review of: Language-Agnostic Suicidal Risk Detection Using Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6RVQTBIN}},
  note         = {Machine review of arXiv:2505.20109}
}
read the original abstract

Suicidal risk detection in adolescents is a critical challenge, yet existing methods rely on language-specific models, limiting scalability and generalization. This study introduces a novel language-agnostic framework for suicidal risk assessment with large language models (LLMs). We generate Chinese transcripts from speech using an ASR model and then employ LLMs with prompt-based queries to extract suicidal risk-related features from these transcripts. The extracted features are retained in both Chinese and English to enable cross-linguistic analysis and then used to fine-tune corresponding pretrained language models independently. Experimental results show that our method achieves performance comparable to direct fine-tuning with ASR results or to models trained solely on Chinese suicidal risk-related features, demonstrating its potential to overcome language constraints and improve the robustness of suicidal risk assessment.

Figures

Figures reproduced from arXiv: 2505.20109 by the authors.

Figure 1
Figure 1. Illustration of the proposed language-agnostic suicidal risk detection framework using LLMs. Speech data is processed through a pretrained speech model, while an ASR system converts speech to text. LLMs extract suicidal risk-related features in Chinese and English, which are used to fine-tune respective language models. Finally, multimodal fusion combines speech and text representa￾tions for suicidal risk prediction… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages

  1. [1]

    According to statistics from the World Health Orga- nization, suicide is the fourth leading cause of death among in- dividuals aged 15–19, affecting both genders [1]

    Introduction Adolescent suicide has emerged as a critical social issue in mod- ern society. According to statistics from the World Health Orga- nization, suicide is the fourth leading cause of death among in- dividuals aged 15–19, affecting both genders [1]. In the United States, the suicide rate among individuals aged 10–24 increased from 6.8 per 100,000...

    work page 2007

  2. [2]

    Language-Agnostic Suicidal Risk Detection Using Large Language Models

    Preliminaries 2.1. Dataset Description In this work, we used the 1st SpeechWellness Challenge (SW1) dataset [20], which was specifically designed for detect- ing suicidal risk among adolescents through speech analysis. This dataset consists of speech recordings from 600 Chinese teenagers aged 10–18 years, including 420 females and 180 males, with an equal...

    work page Pith review arXiv 2025

  3. [3]

    A high-level architecture of our approach is shown in Figure 1

    Method Our proposed language-agnostic suicidal risk detection frame- work integrates both speech and text modalities to boost the robustness and generalizability of suicidal risk assessment. A high-level architecture of our approach is shown in Figure 1. 3.1. Text Modality 3.1.1. Text Data Acquisition using ASR The SW1 Challenge dataset is valuable for su...

    work page

  4. [4]

    Text Modality The experimental results in Table 3 demonstrate the dev set performance of various pretrained language models on the sui- cidal risk assessment on ER and ED tasks

    Results 4.1. Text Modality The experimental results in Table 3 demonstrate the dev set performance of various pretrained language models on the sui- cidal risk assessment on ER and ED tasks. For the Chinese models on the ER task, while directly fine-tuning the Chinese- BERT model on ASR results ( CBERT) achieved Acc (57%) and F1 (64.46%), the GPT-CBERT mo...

    work page

  5. [5]

    Limitation Statement The findings presented in this study are based on the scoring framework of the MINI-KID scale, which assesses current sui- cide risk as at risk or no risk

    Discussion 5.1. Limitation Statement The findings presented in this study are based on the scoring framework of the MINI-KID scale, which assesses current sui- cide risk as at risk or no risk. This classification reflects partic- ipants’ immediate responses to the MINI-KID assessment and should not be interpreted as a prediction of future suicidal be- hav...

    work page

  6. [6]

    Conclusion This study introduced a language-agnostic suicidal risk detec- tion framework leveraging LLMs. Our method achieved perfor- mance comparable to or better than direct fine-tuning with ASR results or monolingual models trained exclusively on Chinese suicidal risk-related features. However, it exhibited a signifi- cant performance gap between the d...

    work page

  7. [7]

    Acknowledgement This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Indus- try Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HI22C1962), and by Brian Impact Foundation, a non-profit or- ganization dedicated to the advancement of science and...

    work page

  8. [8]

    Opi@lt-edi-acl2022: De- tecting signs of depression from social media text using roberta pre-trained language models,

    R. Poswiata and M. Perelkiewicz, “Opi@lt-edi-acl2022: De- tecting signs of depression from social media text using roberta pre-trained language models,” in LTEDI, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:248780097

    work page 2022

Show all 34 references
  1. [9]

    Suicide worldwide in 2019: global health estimates,

    W. H. Organization et al. , “Suicide worldwide in 2019: global health estimates,” 2021

    2019

  2. [10]

    Suicide and homicide death rates among youth and young adults aged 10–24: United states, 2001– 2021,

    S. C. Curtin and M. F. Garnett, “Suicide and homicide death rates among youth and young adults aged 10–24: United states, 2001– 2021,” 2023

    2001

  3. [11]

    Temporal trend in suicide mortality for chinese adolescents, 2008 to 2021,

    Y . Tian, M. Luan, H. Chen, and J. Zhang, “Temporal trend in suicide mortality for chinese adolescents, 2008 to 2021,” JAMA pediatrics, vol. 177, no. 11, pp. 1224–1226, 2023

    2008

  4. [12]

    A review of suicide risk assessment tools and their measured psychometric properties in korea,

    I.-C. Baek, S. Jo, E. J. Kim, G. ryoung Lee, D. H. Lee, and H. J. Jeon, “A review of suicide risk assessment tools and their measured psychometric properties in korea,” Frontiers in Psychiatry , vol. 12, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:235495409

    2021

  5. [13]

    Suicide risk assessment tools, predictive validity findings and utility today: time for a revamp?

    L. E. Roos, J. Sareen, and J. M. Bolton, “Suicide risk assessment tools, predictive validity findings and utility today: time for a revamp?” 2013. [Online]. Available: https: //api.semanticscholar.org/CorpusID:39520759

    2013

  6. [14]

    Factors in the assessment of suicidality in youth,

    K. Posner, G. A. Melvin, B. Stanley, M. A. Oquendo, and M. Gould, “Factors in the assessment of suicidality in youth,”CNS spectrums, vol. 12, no. 2, pp. 156–162, 2007

    2007

  7. [15]

    Automated assessment of psychiatric disorders using speech: A systematic review,

    D. M. Low, K. H. Bentley, and S. S. Ghosh, “Automated assessment of psychiatric disorders using speech: A systematic review,” Laryngoscope Investigative Otolaryngology , vol. 5, pp. 96 – 116, 2019. [Online]. Available: https://api.semanticscholar. org/CorpusID:211829165

    2019

  8. [16]

    Artificial intelligence for suicide assessment using audiovisual cues: a review,

    S. Dhelim, L. L. Chen, H. Ning, and C. D. Nugent, “Artificial intelligence for suicide assessment using audiovisual cues: a review,” Artificial Intelligence Review , vol. 56, pp. 5591– 5618, 2022. [Online]. Available: https://api.semanticscholar.org/ CorpusID:246240326

    2022

  9. [17]

    Exploring the capabilities of a language model- only approach for depression detection in text data,

    M. Sadeghi, B. Egger, R. Agahi, R. Richer, K. Capito, L. H. Rupp, L. Schindler-Gmelch, M. Berking, and B. M. Eskofier, “Exploring the capabilities of a language model- only approach for depression detection in text data,” 2023 IEEE EMBS International Conference on Biomedical a...

    2023

  10. [18]

    Harnessing multimodal approaches for depression detection using large language models and facial expressions,

    M. Sadeghi, R. Richer, B. Egger, L. Schindler-Gmelch, L. H. Rupp, F. Rahimi, M. Berking, and B. M. Eskofier, “Harnessing multimodal approaches for depression detection using large language models and facial expressions,” NPJ Mental Health Research , vol. 3, 2024. [Online]. Ava...

    2024

  11. [19]

    Depression diagnosis: Eeg- based cognitive biomarkers and machine learning,

    K. Boby and S. Veerasingam, “Depression diagnosis: Eeg- based cognitive biomarkers and machine learning,” Behavioural Brain Research , vol. 478, 2024. [Online]. Available: https: //api.semanticscholar.org/CorpusID:273852531

    2024

  12. [20]

    Depression detection and diagnosis based on electroencephalogram (eeg) analysis: A systematic review,

    K. Elnaggar, M. M. El-gayar, and M. M. Elmogy, “Depression detection and diagnosis based on electroencephalogram (eeg) analysis: A systematic review,” Diagnostics, vol. 15, 2025. [Online]. Available: https://api.semanticscholar.org/CorpusID: 275664860

    2025

  13. [21]

    Read speech voice quality and disfluency in individuals with recent suicidal ideation or suicide attempt,

    B. Stasak, J. Epps, H. T. Schatten, I. W. Miller, E. M. Provost, and M. F. Armey, “Read speech voice quality and disfluency in individuals with recent suicidal ideation or suicide attempt,” Speech Commun., vol. 132, pp. 10–20, 2021. [Online]. Available: https://api.semanticsch...

    2021

  14. [22]

    Investigating the speech characteristics of suicidal adolescents,

    S. Scherer, J. P. Pestian, and L.-P. Morency, “Investigating the speech characteristics of suicidal adolescents,” 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 709–713, 2013. [Online]. Available: https: //api.semanticscholar.org/CorpusID:11966138

    2013

  15. [23]

    Adolescent suicidal risk assessment in clinician- patient interaction,

    V . Venek, S. Scherer, L.-P. Morency, A. A. Rizzo, and J. P. Pestian, “Adolescent suicidal risk assessment in clinician- patient interaction,” IEEE Transactions on Affective Computing , vol. 8, pp. 204–215, 2017. [Online]. Available: https: //api.semanticscholar.org/CorpusID:37738621

    2017

  16. [24]

    Unsupervised cross-lingual representation learning for speech recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised cross-lingual representation learning for speech recognition,” Interspeech 2021, 2021

    2021

  17. [25]

    Spontaneous speech-based sui- cide risk detection using whisper and large language mod- els,

    Z. Cui, C. Lei, W. Wu, Y . Duan, D. Qu, J. Wu, R. Chen, and C. Zhang, “Spontaneous speech-based sui- cide risk detection using whisper and large language mod- els,” ArXiv, vol. abs/2406.03882, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:270286119

    2024 arXiv

  18. [26]

    Sever- ity prediction in mental health: Llm-based creation, analysis, evaluation of a novel multilingual dataset,

    K. Skianis, J. Pavlopoulos, and A. S. Do ˘gru¨oz, “Sever- ity prediction in mental health: Llm-based creation, analysis, evaluation of a novel multilingual dataset,” ArXiv, vol. abs/2409.17397, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:272911438

    2024

  19. [27]

    Deep learning and large language models for audio and text analysis in predicting suicidal acts in chinese psychological support hotlines,

    Y . Chen, J. Li, C. Song, Q. Zhao, Y . Tong, and G. Fu, “Deep learning and large language models for audio and text analysis in predicting suicidal acts in chinese psychological support hotlines,” ArXiv, vol. abs/2409.06164, 2024. [Online]. Available: https://api.semanticschol...

    2024 arXiv

  20. [28]

    The 1st speechwellness challenge: Detecting sui- cidal risk among adolescents,

    W. Wu, Z. Cui, C. Lei, Y . Duan, D. Qu, J. Wu, B. Zhou, R. Chen, and C. Zhang, “The 1st speechwellness challenge: Detecting sui- cidal risk among adolescents,” arXiv preprint arXiv:2501.06474, 2025

    2025

  21. [29]

    Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,

    Z. Gao, S. Zhang, I. McLoughlin, and Z. Yan, “Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,” in Interspeech 2022, 2022, pp. 2063– 2067

    2022

  22. [30]

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

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  23. [31]

    Uer: An open-source toolkit for pre-training models,

    Z. Zhao, H. Chen, J. Zhang, W. X. Zhao, T. Liu, W. Lu, X. Chen, H. Deng, Q. Ju, and X. Du, “Uer: An open-source toolkit for pre-training models,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conferenc...

    2019

  24. [33]

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

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

    2022

  25. [34]

    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

  26. [2021]

    Despite its significance, assessing suicidal risk remains largely subjective due to the lack of a definitive characteriza- tion of at-risk individuals

    Notably, adolescents aged 10 to 14 years exhibited an average annual percentage increase of 4.40% [3]. Despite its significance, assessing suicidal risk remains largely subjective due to the lack of a definitive characteriza- tion of at-risk individuals. Various tools have bee...

Pith tools

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