Pith. sign in

REVIEW 4 major objections 6 minor 45 references

Exploring Audio Editing Features as User-Centric Privacy Defenses Against Large Language Model(LLM) Based Emotion Inference Attacks

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Pitch and tempo edits to a voice recording make the evaluated emotion-inference models—GPT-4o and a purpose-built 1D CNN—consistently misclassify the speaker's emotion, offering a familiar, user-friendly privacy defense.

desk verdict Pitch/tempo editing as a usable privacy defense is a genuinely new idea, but the evaluation lacks baselines and aggregates, so the central claim isn't supported. read the letter →

arxiv 2501.18727 v2 pith:2G5OMJDZ submitted 2025-01-30 cs.CR cs.AIcs.LGcs.SDeess.AS

classification cs.CRcs.AIcs.LGcs.SDeess.AS
keywords audioprivacyemotioninferenceattackpitchmanipulationtempospeechrecognitionuser-centriclargelanguagemodelsobfuscation
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 tries to establish a simple user-facing privacy defense: if a person changes the pitch and tempo of their recorded speech using ordinary audio-editing app controls, emotion-inference attackers can no longer correctly identify the speaker's emotion. The authors argue this matters because existing defenses rely on heavy machine-learning pipelines that ordinary users cannot operate, whereas pitch and tempo sliders are already familiar to billions of users. They evaluate the idea on three standard emotional-speech datasets with randomized and structured pitch/tempo combinations, reporting that a trained 1D CNN and GPT-4o consistently misclassify the modified audio. The paper itself hedges the result: reversibility testing is deferred to future work, and the attacker set is limited to those two models. If the claim holds, everyday audio editing becomes a low-cost privacy mechanism for virtual assistants, video calls, and voice recordings.

What carries the argument

The mechanism is the prosodic transformation produced by pitch and tempo manipulation: changing fundamental frequency and speaking rate alters the acoustic cues that speech emotion recognition models rely on, shifting the predicted emotion category. The paper uses these edits as a privacy transformation applied before audio is transmitted to any service, with the user choosing the settings in a familiar audio-editing app. The evaluation apparatus consists of a custom 1D CNN trained on the three datasets, GPT-4o as a multimodal LLM attacker, and two protocols: random pitch/tempo values and a structured grid of pitch values (0, +4, -4, +8, and -8) crossed with tempo values (60, 80, 100, 120, and 140) for systematic analysis.

What would settle it

Re-run the paper's experiments with an unmodified-audio baseline and a larger number of speakers from the same datasets; if the 1D CNN and GPT-4o already misclassify the originals at the same rate as the modified clips, the apparent privacy protection is just baseline inaccuracy. Alternatively, train an attacker on pitch/tempo-modified clips paired with original labels; if it recovers the true emotions at high accuracy, the defense is not robust to adaptation.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that pitch and tempo manipulation effectively obfuscates emotional data in speech, so that an adversary with access to a lightweight 1D CNN or to GPT-4o cannot infer the original emotion from the modified recording. Across RAVDESS, CREMA-D, and TESS, the paper reports that both randomized and systematically varied pitch/tempo combinations shift the predicted emotion away from the ground truth, and that female actors from different datasets show similar shift patterns, such as anger and disgust moving toward fear at high pitch and high tempo. The paper frames this as a demonstration that intuitive, widely available audio-editing features can serve as a user-centric privacy defense, with design goals of lightweight on-device implementation, usability, adaptability, and robustness. It does not yet test the reversibility of the modifications, which its own threat model names as an attack vector.

Load-bearing premise

The claim breaks if the two tested models, evaluated on only a handful of actors without a baseline on unchanged audio, are not a fair stand-in for real emotion-inference adversaries; the paper's own Section 6 also leaves the reversibility attack untested.

Editorial extensions

If this is right

  • A user can apply a few slider adjustments to their own voice recording before sharing it, reducing the risk that an emotion-inference system extracts affective state without consent.
  • The defense requires no training on the user side, only an audio-editing capability already present in common apps, so it can plausibly run on smartphones, smart glasses, and other constrained devices.
  • The structured results give a rough mapping from pitch/tempo settings to emotion shifts, which users or app designers could use to push sensitive emotions into unrelated categories.
  • Because the manipulation happens on the user's device before data reaches a service, the defense is platform-agnostic and does not depend on the assistant or conferencing provider changing its behavior.

Reading between the lines

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

  • An adaptive adversary who knows the defense could retrain the same 1D CNN or fine-tune an LLM on pitch/tempo-modified audio paired with the original labels; the paper does not test this, so its long-term robustness is unknown.
  • The paper's own Section 6 defers reversibility testing; if simple signal processing can undo pitch and tempo shifts, the protection may amount to security by obscurity rather than a hard guarantee.
  • The consistent emotion-shift patterns observed for female actors across datasets hint that the transformation is predictable, and that predictability could let an attacker invert the mapping to guess the original emotion.
  • The approach could be combined with other lightweight transforms, such as formant shifting or time stretching, to make reversal harder, but that combination is not evaluated here.
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 using pitch and tempo manipulation, which are common features in consumer audio editing apps, as a user-centric privacy defense against emotion inference attacks from speech. It describes a threat model that includes DNN-based and LLM-based adversaries, trains a 1D-CNN, and evaluates the defense using that CNN and GPT-4o on three datasets (RAVDESS, CREMA-D, TESS). The central claim, stated in the abstract and Section 6, is that modifying pitch and tempo causes the evaluated attack models to consistently fail to infer emotions, thereby protecting users' emotional privacy. The paper also sketches design principles for lightweight on-device implementation.

Significance. The core idea is novel and practically motivated: leveraging familiar audio editing features could make privacy protection more accessible than prior ML-heavy obfuscation methods. However, the current evidence does not support the central claim. The evaluation lacks a baseline on unmodified audio, is limited to a handful of hand-selected examples, and the abstract advertises reversibility testing that is deferred to future work. If the missing baseline experiments were added and the claims were scaled back to a preliminary feasibility study, the paper could be a useful starting point; in its present form, the conclusions outrun the data.

major comments (4)
  1. [§5.1 / §6] The paper never reports the attack models' accuracy on unmodified audio, so the conclusion in Section 6 that 'the evaluated attack models consistently failed to accurately infer emotions from the modified data' is unsupported. Without a baseline, we cannot determine whether pitch/tempo modification reduces inference accuracy at all, and the paper's own tables contain rows where both models predict the original emotion (e.g., Table 3, first row, original Neutral; Table 7, rows for Happy at pitch 0.0, tempo 132 and pitch 8.5, tempo 100). The claim of consistent failure is therefore contradicted by the presented data.
  2. [Abstract / §6] The abstract and Section 1 state that the threat model includes 'reversibility testing' and the abstract says this was rigorously evaluated, but Section 6 lists reversibility as future work ('We will also investigate the reversibility of pitch and tempo modifications') and no reversibility experiment appears in Section 5. The advertised evaluation is absent.
  3. [§5.1] The randomized evaluation selects one male and one female actor per dataset (TESS has only female actors) and presents only hand-selected rows in Tables 3–7, with no description of how rows were chosen, no aggregate accuracy over all modified files, no error bars, and no statistical tests. These tables cannot support the generalizability claims in Section 6 or the demographic-bias analysis promised in the Introduction.
  4. [§5.1] The observation that female actors 'exhibit notable commonalities in how their emotions are affected' is based on two female actors from different datasets and is not quantified; it is an anecdotal pattern, not a finding, and the specific claim that increasing tempo while keeping pitch constant changes neutral to disgust is not supported by a consistent pattern in the tables.
minor comments (6)
  1. [Abstract] The phrase 'and and reversibility testing' contains a duplicated 'and'.
  2. [§4.3] The sentence 'trained it using the RAVDESS, CREMA-D and three prominent dataset' should likely read 'these three prominent datasets'; the wording is confusing and the number of datasets is inconsistent with the experiments.
  3. [§5.1] The typo 'REVDESS' should be 'RAVDESS'.
  4. [Tables 3–7] The units for pitch and tempo are not specified; values such as -9.4 and 72 should be labeled (e.g., semitones, BPM) so the reader can interpret the manipulation settings.
  5. [Table 8] The row '0 -8' and the phrase 'Neutral tempo (0 BPM)' are confusing, since a tempo of 0 BPM is not meaningful; please clarify what value 0 represents.
  6. [§4.3] No train/validation/test split, input features, or hyperparameters are reported for the 1D-CNN, so the claimed 97% test accuracy cannot be reproduced or assessed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the defense claim is empirical and not reduced to its inputs; limitations are correctness concerns, not circularity.

full rationale

The paper's central claim is empirical: pitch and tempo modifications make the evaluated 1-D CNN and GPT-4o fail to infer the original emotion labels. It is not a derivation in which an output quantity is defined in terms of an input quantity. No parameter is fitted to a subset of the data and then renamed as a prediction: the 1-D CNN is trained on unmodified benchmark audio and later evaluated on modified audio, while GPT-4o is used as a pre-trained model, so the observed failures are not forced by construction. The only co-authored citation, reference [9] by Shahid and Imteaj, supports background motivation about attribute inference in video-conferencing settings; it is not load-bearing for the pitch/tempo privacy claim. The paper does not invoke a uniqueness theorem from the same authors, and no known result is merely renamed. Weaknesses such as the absence of baseline accuracy on unmodified audio, the hand-selected table rows, the lack of aggregate statistics, and the deferred reversibility testing are threats to the validity or strength of the empirical conclusion, but they are not circularity: the claim is not defined in terms of the attack model's outputs, nor is the success metric a renamed training objective. No specific circular step can be quoted from the paper, so the appropriate finding is no significant circularity.

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

The report has no mathematical derivation; the central claim rests on experimental assumptions. The main assumptions are that emotion inference is a realistic threat, that pitch and tempo edits shift classifier outputs at the chosen grid values, that the two chosen attackers are representative, and that usability survives the edits. None of these are quantified against external baselines.

free parameters (1)
  • Pitch/tempo manipulation grid = pitch: -8, -4, 0, 4, 8; tempo: 60, 80, 100, 120, 140 (random values also used)
    Chosen by hand as the experimental conditions; the privacy effect is measured only at these settings, so the claim depends on them. Not fitted to data, but they define the scope of the result.
assumptions (5)
  • domain assumption Emotion inference from audio is a meaningful privacy threat in the target systems.
    Motivates the whole defense; taken from prior literature [5][6] and not independently established here.
  • domain assumption Pitch and tempo changes alter perceived or prosodic emotion sufficiently to shift classifier outputs.
    Core mechanism; supported by cited music and emotion literature [24][25], but for speech SER it is assumed and then observed in selected tables.
  • domain assumption GPT-4o can analyze audio files and infer emotion as an adversary.
    Used as a black-box attacker; the paper does not describe how GPT-4o was prompted or how audio was fed to it.
  • ad hoc to paper The selected actors (one male, one female per dataset) are representative enough to generalize across populations.
    Selection is not justified and TESS has only female actors; generalization is asserted.
  • ad hoc to paper Modified audio remains usable for downstream tasks.
    Usability is a design goal, but no intelligibility, listening, or task-utility measurement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Audio Editing Features as User-Centric Privacy Defenses Against Large Language Model(LLM) Based Emotion Inference Attacks." pith.science (2026). https://pith.science/paper/2G5OMJDZ

@misc{pith2026250118727,
  author       = {Pith},
  title        = {Pith review of: Exploring Audio Editing Features as User-Centric Privacy Defenses Against Large Language Model(LLM) Based Emotion Inference Attacks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2G5OMJDZ}},
  note         = {Machine review of arXiv:2501.18727}
}
read the original abstract

The rapid proliferation of speech-enabled technologies, including virtual assistants, video conferencing platforms, and wearable devices, has raised significant privacy concerns, particularly regarding the inference of sensitive emotional information from audio data. Existing privacy-preserving methods often compromise usability and security, limiting their adoption in practical scenarios. This paper introduces a novel, user-centric approach that leverages familiar audio editing techniques, specifically pitch and tempo manipulation, to protect emotional privacy without sacrificing usability. By analyzing popular audio editing applications on Android and iOS platforms, we identified these features as both widely available and usable. We rigorously evaluated their effectiveness against a threat model, considering adversarial attacks from diverse sources, including Deep Neural Networks (DNNs), Large Language Models (LLMs), and and reversibility testing. Our experiments, conducted on three distinct datasets, demonstrate that pitch and tempo manipulation effectively obfuscates emotional data. Additionally, we explore the design principles for lightweight, on-device implementation to ensure broad applicability across various devices and platforms.

Figures

Figures reproduced from arXiv: 2501.18727 by the authors.

Figure 1
Figure 1. Interface of audio editing apps: Up Tempo, Au [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The Privacy-Preserving mechanism integrated system model. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. 1D-CNN Model Architecture 4.3 Trained Threat Models As our potential attacker possesses the capability of using deep learning model including immensely powerful LLM models for the aforementioned inference attack, for attack simulation first we designed a 1-D CNN model from scratch and trained it using the RAVDESS, CREMA-D and three prominent dataset for emotion recognition. The reason for using 1-D CNN is that not o… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [1]

    Music enhances activity in the hypothalamus, brainstem, and anterior cerebellum during script-driven imagery of affective scenes

    Chia-Wei Li, Tzu-Han Cheng, and Chen-Gia Tsai. Music enhances activity in the hypothalamus, brainstem, and anterior cerebellum during script-driven imagery of affective scenes. Neuropsychologia, 133:107073, 2019. 9 The 6th AAAI Workshop on Privacy-Preserving Artificial Intelligence PPAI-25

  2. [2]

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

    Matthew B Hoy. Alexa, siri, cortana, and more: an introduction to voice assistants. Medical reference services quarterly, 37(1):81–88, 2018

  3. [3]

    3-d sound for virtual reality and multimedia

    Durand R Begault and Leonard J Trejo. 3-d sound for virtual reality and multimedia. Technical report, 2000

  4. [4]

    Audio-visual perception of omnidirectional video for virtual reality applications

    Fang-Yi Chao, Cagri Ozcinar, Chen Wang, Emin Zerman, Lu Zhang, Wassim Hamidouche, Olivier Deforges, and Aljosa Smolic. Audio-visual perception of omnidirectional video for virtual reality applications. In 2020 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), pages 1–6, 2020

  5. [5]

    New challenges for content privacy in speech and audio

    Jennifer Williams, Karla Pizzi, Shuvayanti Das, and Paul-Gauthier Noé. New challenges for content privacy in speech and audio. arXiv preprint arXiv:2301.08925, 2023

  6. [6]

    Privacy risks in speech emotion recognition: A systematic study on gender inference attack

    Basmah Alsenani, Tanaya Guha, and Alessandro Vinciarelli. Privacy risks in speech emotion recognition: A systematic study on gender inference attack. 2023

  7. [7]

    Attribute Inference Attack of Speech Emotion Recognition in Federated Learning Settings

    Tiantian Feng, Hanieh Hashemi, Rajat Hebbar, Murali Annavaram, and Shrikanth S Narayanan. Attribute inference attack of speech emotion recognition in federated learning settings. arXiv preprint arXiv:2112.13416, 2021

  8. [8]

    Zooming into video conferencing privacy

    Dima Kagan, Galit Fuhrmann Alpert, and Michael Fire. Zooming into video conferencing privacy. IEEE Transactions on Computational Social Systems, 11(1):933–944, 2024

Show all 45 references
  1. [9]

    Shahid and Ahmed Imteaj

    Abdur R. Shahid and Ahmed Imteaj. Securing User Privacy in Cloud-Based Whiteboard Services Against Health Attribute Inference Attacks . IEEE Transactions on Artificial Intelligence, 5(08):3872–3885, August 2024

  2. [10]

    Sok: Data privacy in virtual reality

    Gonzalo Munilla Garrido, Vivek Nair, and Dawn Song. Sok: Data privacy in virtual reality. Proceedings on Privacy Enhancing Technologies, 2024(1):21–40, January 2024

  3. [11]

    Privacy preservation in artificial intelligence and extended reality (ai-xr) metaverses: A survey, 2023

    Mahdi Alkaeed, Adnan Qayyum, and Junaid Qadir. Privacy preservation in artificial intelligence and extended reality (ai-xr) metaverses: A survey, 2023

  4. [12]

    Privacy of the metaverse: Current issues, ai attacks, and possible solutions

    Chamara Sandeepa, Shen Wang, and Madhusanka Liyanage. Privacy of the metaverse: Current issues, ai attacks, and possible solutions. In 2023 IEEE International Conference on Metaverse Computing, Networking and Applications (MetaCom), pages 234–241, 2023

  5. [13]

    Gradient-level differential privacy against attribute inference attack for speech emotion recognition

    Haijiao Chen, Huan Zhao, and Zixing Zhang. Gradient-level differential privacy against attribute inference attack for speech emotion recognition. IEEE Signal Processing Letters, 31:3124–3128, 2024

  6. [14]

    Privacy-preserving voice analysis via disentangled representa- tions

    Ranya Aloufi, Hamed Haddadi, and David Boyle. Privacy-preserving voice analysis via disentangled representa- tions. In Proceedings of the 2020 ACM SIGSAC Conference on Cloud Computing Security Workshop, CCSW’20, page 1–14, New York, NY , USA, 2020. Association for Computing Machinery

  7. [15]

    An opt-in framework for privacy protection in audio-based applications

    Wei-Cheng Wang, Sander De Coninck, Sam Leroux, and Pieter Simoens. An opt-in framework for privacy protection in audio-based applications. IEEE Pervasive Computing, 21(4):17–24, 2022

  8. [16]

    Privacy against real-time speech emotion detection via acoustic adversarial evasion of machine learning

    Brian Testa, Yi Xiao, Harshit Sharma, Avery Gump, and Asif Salekin. Privacy against real-time speech emotion detection via acoustic adversarial evasion of machine learning. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., 7(3), September 2023

  9. [17]

    Apple intelligence

    Apple Inc. Apple intelligence. https://www.apple.com/apple-intelligence/, 2024. Accessed: 2024-11- 28

  10. [18]

    Google play store

    Google. Google play store. https://play.google.com/, 2024. Accessed: 2024-11-28

  11. [19]

    Apple app store

    Apple Inc. Apple app store. https://www.apple.com/app-store/, 2024. Accessed: 2024-11-28

  12. [20]

    Android official website

    Google. Android official website. https://www.android.com/, 2024. Accessed: 2024-11-28

  13. [21]

    ios 18 features

    Apple Inc. ios 18 features. https://www.apple.com/ios/ios-18/, 2024. Accessed: 2024-11-28

  14. [22]

    User-level differential privacy against attribute inference attack of speech emotion recognition in federated learning

    Tiantian Feng, Raghuveer Peri, and Shrikanth Narayanan. User-level differential privacy against attribute inference attack of speech emotion recognition in federated learning. arXiv preprint arXiv:2204.02500, 2022

  15. [23]

    Privacy-enhanced federated learning against attribute inference attack for speech emotion recognition

    Huan Zhao, Haijiao Chen, Yufeng Xiao, and Zixing Zhang. Privacy-enhanced federated learning against attribute inference attack for speech emotion recognition. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

  16. [24]

    Effects of pitch and musical sounds on body-representations when moving with sound

    Judith Ley-Flores, Eslam Alshami, Aneesha Singh, Frédéric Bevilacqua, Nadia Bianchi-Berthouze, Ophelia Deroy, and Ana Tajadura-Jiménez. Effects of pitch and musical sounds on body-representations when moving with sound. Scientific reports, 12(1):2676, 2022

  17. [25]

    Influence of tempo and rhythmic unit in musical emotion regulation

    Alicia Fernández-Sotos, Antonio Fernández-Caballero, and José M Latorre. Influence of tempo and rhythmic unit in musical emotion regulation. Frontiers in computational neuroscience, 10:80, 2016. 10 The 6th AAAI Workshop on Privacy-Preserving Artificial Intelligence PPAI-25

  18. [26]

    {SMACK}: Semantically meaningful adversarial audio attack

    Zhiyuan Yu, Yuanhaur Chang, Ning Zhang, and Chaowei Xiao. {SMACK}: Semantically meaningful adversarial audio attack. In 32nd USENIX Security Symposium (USENIX Security 23), pages 3799–3816, 2023

  19. [27]

    A cross-domain exploration of audio and textual data for multi-modal emotion detection

    Mohd Ariful Haque, Roy George, Rakib Hossain Rifat, Md Shihab Uddin, Marufa Kamal, and Kishor Datta Gupta. A cross-domain exploration of audio and textual data for multi-modal emotion detection. In Proceedings of the 17th International Conference on PErvasive Technologies Rela...

  20. [28]

    Theresa Goetz, Pulkit Arora, F. X. Erick, Nina Holzer, and Shrutika Sawant. Self-supervised representation learning using multimodal transformer for emotion recognition. In Proceedings of the 8th International Workshop on Sensor-Based Activity Recognition and Artificial Intell...

  21. [29]

    Speech emotion recognition based on 1d cnn and mfcc

    Gaoyun Li, Yong Liu, and Xiong Wang. Speech emotion recognition based on 1d cnn and mfcc. In2023 IEEE 5th International Conference on Civil Aviation Safety and Information Technology (ICCASIT), pages 956–960, 2023

  22. [30]

    Emotion recognition in audio and video using deep neural networks

    Mandeep Singh and Yuan Fang. Emotion recognition in audio and video using deep neural networks. arXiv preprint arXiv:2006.08129, 2020

  23. [31]

    Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances

    Zehui Wu, Ziwei Gong, Lin Ai, Pengyuan Shi, Kaan Donbekci, and Julia Hirschberg. Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances. arXiv preprint arXiv:2407.21315, 2024

  24. [32]

    Secap: Speech emotion captioning with large language model

    Yaoxun Xu, Hangting Chen, Jianwei Yu, Qiaochu Huang, Zhiyong Wu, Shi-Xiong Zhang, Guangzhi Li, Yi Luo, and Rongzhi Gu. Secap: Speech emotion captioning with large language model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19323–19331, 2024

  25. [33]

    Alexa official website

    Amazon. Alexa official website. https://www.alexa.com/, 2024. Accessed: 2024-11-28

  26. [34]

    Siri by apple

    Apple Inc. Siri by apple. https://www.apple.com/siri/, 2024. Accessed: 2024-11-28

  27. [35]

    Google assistant official website

    Google. Google assistant official website. https://assistant.google.com/, 2024. Accessed: 2024-11-28

  28. [36]

    Chatgpt official website

    OpenAI. Chatgpt official website. https://chatgpt.com/, 2024. Accessed: 2024-11-28

  29. [37]

    Gemini by google

    Google. Gemini by google. https://gemini.google.com/, 2024. Accessed: 2024-11-28

  30. [38]

    Google meet

    Google. Google meet. https://meet.google.com/, 2024. Accessed: 2024-11-28

  31. [39]

    Zoom video communications

    Zoom Video Communications, Inc. Zoom video communications. https://www.zoom.com/, 2024. Accessed: 2024-11-28

  32. [40]

    Meta smart glasses

    Meta Platforms, Inc. Meta smart glasses. https://www.meta.com/smart-glasses/, 2024. Accessed: 2024- 11-28

  33. [41]

    The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english

    Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018

  34. [42]

    Cooper, Michael K

    Houwei Cao, David G. Cooper, Michael K. Keutmann, Ruben C. Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE Transactions on Affective Computing, 5(4):377–390, 2014

  35. [43]

    Kathleen Pichora-Fuller and Kate Dupuis

    M. Kathleen Pichora-Fuller and Kate Dupuis. Toronto emotional speech set (TESS), 2020

  36. [44]

    emotion2vec: Self-supervised pre-training for speech emotion representation, 2023

    Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation, 2023

  37. [45]

    wav2vec 2.0: A framework for self-supervised learning of speech representations, 2020

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations, 2020. 11 The 6th AAAI Workshop on Privacy-Preserving Artificial Intelligence

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.