Robust Device-Directed Speech Detection on the Edge via Adversarial Preprocessing and Semantic Masking

View at publisher Summarise with VIRAT
Authors: Hanyu Deng, Siyu Huang, Shijie Cao, Haishan Chen, Haoran Deng, Junying Yuan
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Robust Device-Directed Speech (DDS) detection on edge devices is a critical enabler for privacy-preserving, wake-word-free voice interaction. However, due to strict privacy constraints, large-scale data collection is often prohibitive. This data scarcity leads to a critical pathology known as “shortcut learning,” where lightweight models overfit to non-semantic artifacts — such as volume variance or silence duration — rather than learning intrinsic prosody. To mitigate this issue without increasing model complexity, we propose a data-centric “Dual-Defense” framework. First, we introduce an Adversarial Data Preprocessing Pipeline (ADPP) as a physical defense to systematically eliminate signal-level shortcuts, including proximity effects and channel biases. Second, we employ a Semantic Masking Strategy as a semantic defense, which acts as a regularization mechanism to decouple prosodic intent from phonetic content during training. Experimental results on a custom privacy-sensitive dataset demonstrate that our approach significantly enhances the robustness of a lightweight Convolutional Recurrent Neural Network (CRNN) (0.32M parameters). The proposed framework boosts the F1-score from 0.84 to 0.94 and reduces false positives by approximately 85% compared to the baseline, proving its efficacy for low-resource edge deployments.

Fields of Study & Subjects

Speech Recognition and Synthesis Speech and Audio Processing Speech and dialogue systems

Publication details

Journal / source
International Conference on Artificial Intelligence and Social Network Systems (AISNS)
Pages
101–106
Publication date
9 January 2026
Language
EN