Submission Types
Long papers may include up to 8 pages of main content. Short papers may include up to 4 pages of main content. References and other allowed sections should follow ARR/ACL submission requirements.
Call for Papers
Secure, safe, and privacy-preserving speech and audio AI
We invite original work on risks, evaluations, defenses, and deployment practices for trustworthy speech and audio models.
Scope
This workshop focuses on security, safety, and privacy risks in speech and audio models across the full lifecycle, including attacks and defenses, robustness and safety evaluation, privacy leakage, and mitigation strategies.
We also welcome system-level work on evaluation protocols, release practices, privacy-preserving deployment, and integration into interactive, agent-based, and multimodal systems.
Submission Details
TrustAudio accepts long papers of up to 8 pages and short papers of up to 4 pages. Submissions must be anonymized, follow ARR/ACL formatting, and be submitted through OpenReview. ARR commitments should use the OpenReview ARR Commitment site.
Topics of Interest
- Digital adversarial attacks on speech and audio models, including ASR, TTS, and speaker recognition.
- Physical-world and real-world audio attacks, including over-the-air and replay attacks.
- Robustness, safety, and reliability evaluation of audio-language and multimodal models, benchmarks, and protocols.
- Deepfake and synthetic voice misuse, including voice cloning and impersonation.
- Hallucinations, misalignment, and other unintended behaviors in audio and multimodal generative models.
- Data poisoning and backdoor attacks across the model lifecycle, including data, training, fine-tuning, and release.
- Privacy risks and privacy-preserving training or deployment for speech and audio systems.
- Model extraction and IP attacks, plus provenance and attribution for synthetic audio.
- Defensive methods and system-level safeguards for speech, audio, and multimodal systems.
- Safety and robustness in multilingual, multicultural, and underrepresented languages.
- Safety evaluation, red-teaming, and alignment of LLMs, audio-language models, and multimodal foundation models.
- Prompt injection, jailbreaks, instruction-following failures, and unsafe tool use in LLMs and multimodal frameworks.
- Uncertainty, calibration, refusal behavior, and responsible generation across language, speech, audio, and multimodal systems.
Review Process
Submissions will be reviewed through a double-blind peer-review process. Papers and supplementary materials must be anonymized for review.
Submission Platform
Submit via OpenReview. ARR-reviewed papers should be committed through the OpenReview ARR Commitment site and will be considered if they comply with the workshop submission and cross-submission policies.
Formatting
Papers should use the ARR/ACL style files and follow ARR submission requirements, including required limitations and applicable ethics sections.
Originality and Cross-Submission
Submissions must describe original, unpublished work. Cross-submissions and dual submissions are not allowed; papers must not be under review, accepted, or committed for publication at another archival venue during the TrustAudio review period.
Presentation
Accepted papers are expected to be presented at the workshop. Additional camera-ready and proceedings instructions will follow AACL-IJCNLP workshop guidance.