Deep Denoising for Hearing Aid Applications

Marc Aubreville, Kai Ehrensperger, Andreas K. Maier, Tobias Rosenkranz, Benjamin Graf, Henning Puder

Hearing aids have come a long way, but anyone who has used one in a crowded restaurant knows they still struggle with "non-stationary" noise—the unpredictable clatter of dishes or the roar of a nearby table. While directional microphones help, they require the speaker to be right in front of the user. In a new paper, researchers from FAU Erlangen-Nürnberg and Sivantos GmbH (the makers of Signia hearing aids) propose a deep learning approach to clean up audio in real-time without the lag that usually plagues complex algorithms.

TL;DR

The researchers developed a deep neural network (DNN) that predicts optimal noise-reduction gains in real-time for hearing aids. By using an "asymmetric context"—looking back at 200ms of past audio but only 2ms into the future—the system outperforms traditional methods in both objective clarity and subjective listening tests while staying within the strict 8ms latency limit required for comfortable hearing aid use.

What Problem Does This Paper Address?

The "Holy Grail" of hearing aid technology is the ability to isolate speech from a noisy background (the "Cocktail Party Problem"). Most current devices use statistical methods that work well for steady hums (like an airplane cabin) but fail when noise changes rapidly.

Deep learning offers a solution, but it faces a massive hurdle in hearing aids: latency. If a hearing aid takes too long to process sound (usually more than 10ms), the processed audio interferes with the natural sound leaking into the ear, creating a distracting "echo" or hollow sound known as the comb-filter effect. Most high-performing DNNs require a "lookahead" into future audio frames to understand context, which introduces too much delay for medical-grade hardware.

Methods

The authors propose a single-channel denoising system built on a fully connected DNN with three hidden layers (2048 nodes each).

The Workflow: 1. Filter Bank: The audio is split into 48 frequency channels using a specialized hearing-aid filter bank. 2. Feature Extraction: The system calculates the log power spectrum and normalizes it. 3. Asymmetric Context: This is the "secret sauce." The network is fed a window of audio that includes a significant amount of "lookback" (past data) but almost zero "lookahead" (future data). This allows the model to learn temporal patterns in speech without adding processing delay. 4. Wiener Gain Prediction: Instead of trying to "reconstruct" clean speech (which can sound robotic), the DNN predicts a "gain vector"—essentially a set of volume knobs for each of the 48 frequency channels—to suppress noise while preserving the original speech phase.

Datasets

The model was trained and tested using:

  • Noise: 49 real-world noise signals recorded across Europe using actual hearing aid microphones.
  • Speech: German sentences from the EUROM database.
  • Augmentation: To ensure robustness, they mixed speech with up to four different noise sources at various Signal-to-Noise Ratios (SNRs) ranging from -10dB to +20dB.

Key Insights and Ideas

  • The 200ms Sweet Spot: The researchers found that the network’s performance improves as you give it more past context, but the benefits plateau at around 200–300ms. This interestingly aligns with the average rate of speech syllables (~4Hz), suggesting the network is learning to recognize the temporal "envelope" of human speech.
  • Future vs. Past: While "looking into the future" helps, the study shows that a tiny lookahead (just 2ms) combined with a long lookback is enough to significantly outperform traditional algorithms.
  • Integration Ready: Because the DNN outputs a gain vector rather than a raw waveform, it can be easily plugged into existing hearing aid processing chains (like automatic volume control) without a total redesign.

Key Results

The researchers compared their DNN against a "state-of-the-art" recursive minimum tracking baseline (a common industry standard).

  • Objective Metrics: Using the STOI (Short-Term Objective Intelligibility) metric, the DNN showed significant improvements in low-SNR (very noisy) conditions, whereas the baseline method actually decreased intelligibility in some cases.
  • Subjective Testing: In a MUSHRA (MUlti Stimulus test with Hidden Reference and Anchor) test with 20 listeners, the DNN-generated audio was consistently rated higher in quality than the baseline across all noise levels. In some high-SNR scenarios, the DNN was nearly indistinguishable from the "ideal" theoretical limit.

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Phase Preservation: Because the system uses a Wiener filter approach, it keeps the "noisy phase" of the original signal. In extremely noisy environments, this can limit the ultimate ceiling of audio quality.
  • Computational Power: While the latency is low, running a 2048-node 3-layer MLP is still computationally expensive for a tiny hearing aid battery. The paper focuses on the algorithmic feasibility; the next step would be optimizing the network architecture (e.g., through pruning or quantization) for low-power silicon.
  • Generalization: The study used a specific set of noises; further testing is needed to see how it handles "unseen" noise types not present in the training set.

Summary updated

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.
BibTeX
@inproceedings{aubreville2018deep,
  author = {Marc Aubreville and Kai Ehrensperger and Andreas K. Maier and Tobias Rosenkranz and Benjamin Graf and Henning Puder},
  title = {{Deep Denoising for Hearing Aid Applications}},
  booktitle = {IWAENC},
  year = {2018},
  url = {https://openreview.net/forum?id=hTkx5zxLAqe},
}

Loading related content...

Figure