DASH: Dual-view self-distillation makes speech recognition robust to noise without clean performance loss

Automatic speech recognition (ASR) systems notoriously degrade in noisy environments, and typical supervised noise-augmented fine-tuning introduces a robustness-clean trade-off while overfitting to specific corruptions. To address this, a team of researchers will present DASH (Dual-View Self-Distillation with Multi-Layer Hidden Representations) at Interspeech 2026. DASH is a self-distillation framework that trains a single model to learn consistency between paired clean and noisy views of speech. It extracts features from multiple encoder layers—capturing everything from low-level acoustics to high-level semantics—and minimizes KL divergence between prototype assignment distributions of the two views, stabilizing training without requiring labels.

Experiments on the LibriSpeech benchmark show that DASH consistently improves recognition accuracy under diverse noisy conditions while fully preserving clean accuracy. The method achieves this robustness via a label-free pre-training stage that adds only about 4% overhead beyond standard fine-tuning—a practical advantage for deployment. By eliminating the robustness-clean trade-off and requiring minimal extra compute, DASH offers a promising path toward production-ready ASR that works reliably in real-world acoustic environments.