Hearing aid DNN separates speech, music, and noise power
A single low-complexity neural network replaces multiple acoustic scene estimators for hearing aids.
A team from academia and industry—Mats Lang, Thomas Haubner, Nina Kiessling, Christoph Hoog Antink, and Henning Puder—developed a Deep Neural Network (DNN) that estimates the power of speech, music, and noise simultaneously across time and frequency. The work, accepted to the International Workshop on Acoustic Signal Enhancement (IWAENC) 2026, targets hearing aid signal processing where current systems run separate estimators for scene classification, Voice Activity Detection (VAD), and signal-to-noise ratio. This fragmented approach is computationally heavy and ignores dependencies between tasks.
The proposed model uses a causal, low-complexity DNN to decompose an acoustic mixture into interpretable components: speech, music, and noise power proportions. From this unified representation, multiple downstream metrics—such as VAD—can be derived with simple post-processing. In experiments, the representation achieved VAD performance comparable to state-of-the-art estimators while outputting a much richer scene description. This suggests hearing aids could adapt more intelligently to complex environments, like busy restaurants or concert halls, using less battery power and providing more natural sound processing. The paper was posted on arXiv (2608.17482) on August 18, 2026.
- Causal low-complexity DNN estimates time- and frequency-dependent speech, music, and noise power proportions
- Matches SOTA Voice Activity Detection performance while providing richer scene representation
- Accepted to IWAENC 2026; reduces computational complexity by replacing multiple independent estimators
Why It Matters
Hearing aid users get faster, more accurate automatic adaptation to complex sound environments with less processing overhead.