Abstract
Reduction of unwanted environmental noises is an important feature of today's hearing aids (HA), which is why noise reduction is nowadays included in almost every commercially available device. The majority of these algorithms, however, is restricted to the reduction of stationary noises. In this work, we propose a denoising approach based on a three hidden layer fully connected deep learning network that aims to predict a Wiener filtering gain with an asymmetric input context, enabling real-time applications with high constraints on signal delay. The approach is employing a hearing instrument-grade filter bank and complies with typical hearing aid demands, such as low latency and on-line processing. It can further be well integrated with other algorithms in an existing HA signal processing chain. We can show on a database of real world noise signals that our algorithm is able to outperform a state of the art baseline approach, both using objective metrics and subject tests.
Introduction
A primary challenge in hearing aid (HA) technology is the "Cocktail Party Problem"—isolating speech from complex, non-stationary background noise. While traditional statistical methods effectively suppress stationary noises like engine hums, they often fail with unpredictable environmental sounds. Deep Neural Networks (DNNs) offer a promising alternative but are constrained by the strict latency requirements of HAs. Processing delays exceeding 10ms can cause the comb-filter effect, where processed sound interferes with natural sound entering the ear. This paper presents a real-time DNN-based denoising system that achieves high performance within these latency bounds.
System Architecture and Methodology
The proposed system utilizes a single-channel denoising approach integrated into a standard HA signal processing chain.
As shown in the architecture diagram, the input signal is decomposed into 48 frequency channels using an Analysis Filter Bank (AFB). Features are extracted by calculating the log power spectrum, followed by normalization. The core of the system is a fully connected DNN with three hidden layers (2048 nodes each).
A key innovation is the use of asymmetric context. The network receives a window of audio with a significant "lookback" (past data) but minimal "lookahead" (future data). This configuration allows the model to capture temporal speech patterns without introducing prohibitive processing delays. The DNN predicts a Wiener gain vector, which is applied to the noisy signal before reconstruction via a Synthesis Filter Bank (SFB).
Experimental Setup
The model was trained and evaluated using:
- Speech: German sentences from the EUROM database.
- Noise: 49 real-world noise recordings from various European environments.
- Augmentation: Speech and noise were mixed at SNRs ranging from -10dB to +20dB, with up to four noise sources per mixture to ensure robustness.
Results and Discussion
Temporal Context Analysis
Investigation into the input context revealed that performance gains from "lookback" data plateau at approximately 200–300ms. This duration corresponds to the typical syllabic rate of human speech (~4Hz), suggesting the network effectively learns the temporal envelope of speech. Furthermore, a minimal lookahead of just 2ms was found to be sufficient for significant performance improvements over traditional baselines.
Objective Performance
The DNN approach was compared against a state-of-the-art recursive minimum tracking baseline.
As illustrated in the performance metrics, the DNN ("predicted gain") consistently outperformed the baseline across multiple metrics. In low-SNR conditions, the DNN significantly improved the Short-Term Objective Intelligibility (STOI) metric, whereas the baseline often degraded it. The DNN also achieved higher Noise Reduction (NR) with comparable or lower Speech Distortion (SD) than the baseline.
Subjective Evaluation
Subjective quality was assessed via a MUSHRA test with 20 listeners. The DNN-processed audio was rated significantly higher than the baseline across all noise levels. In high-SNR scenarios, the DNN's performance approached the theoretical "ideal gain" limit, making it nearly indistinguishable from the reference signal.
Conclusion
This work demonstrates the feasibility of using deep learning for real-time, low-latency speech enhancement in hearing aids. By employing an asymmetric context and predicting gain vectors, the system outperforms traditional industry standards while remaining compatible with existing HA hardware constraints. Future research will focus on optimizing the network for low-power silicon through quantization and exploring phase-aware processing to further enhance audio quality in extreme noise.
Summary updated