Speech enhancement models improve adaptively without retraining
BLINC: Blind Calibration For Training-Free Speech Enhancement Adaptation
Sound
Summary
Speech enhancement tools often perform worse when used in new or noisy environments they haven't seen before. The authors propose BLINC, a method that adjusts the output of these tools on the fly without changing their internal settings or needing extra training. BLINC works by adjusting the predicted sound patterns to better match typical noisy speech using features extracted directly from the noisy audio. This approach improves the sound quality more reliably and with less computational effort compared to other methods that retrain the model during use.
What this means in practice
- •For audio software developers: Integrate BLINC to improve speech enhancement performance in new noise environments without modifying model weights at runtime.
- •For telecommunications engineers: Deploy BLINC in communication devices to maintain clearer speech quality across changing acoustic conditions without extra training overhead.
Authors
Tobias Raichle, Ekaterina Gavrilko, Bin Yang
Abstract
Speech enhancement (SE) models degrade under domain shifts and have to adapt to unseen target domains during deployment. Most existing test-time adaptation (TTA) methods for SE do so by adapting a subset of the model weights using a self-supervised loss, which requires backpropagation at test-time and permanently alters the model. We instead recalibrate the prediction itself and propose BLINC, a training-free TTA method that remaps the predicted time-frequency mask onto a bimodal target distribution by histogram matching. At test-time, the target distribution is parameterized from blind features of the noisy recording, so neither a reference distribution from a classical algorithm nor online metric optimization is involved. BLINC improves the overall quality of both evaluated SE models on almost every target condition and matches or exceeds the loss-based TTA baselines at minimal overhead.