In-the-Loop Training Eliminates Hearing Aid Whistling at High Gains
New deep learning method prevents feedback howling even at extreme amplification levels.
Acoustic feedback—the annoying whistling sound—limits the maximum amplification hearing aids can provide. Traditional methods rely on adaptive filtering, and more recently, open-loop deep neural network feedback cancellation (DFC-OL) has been proposed. However, DFC-OL becomes unstable when gain is pushed high because it never encounters unstable conditions during training.
To solve this, Voit and Doclo introduce in-the-loop trained DFC (DFC-IL). By integrating the feedback cancellation model directly into the optimization loop, the network learns to suppress howling during training itself. A two-stage strategy first pre-trains on stable systems, then fine-tunes on a wider gain range that includes unstable conditions, teaching robust howling reduction.
Results on measured acoustic feedback paths show that at low gains, DFC-IL performs similarly to DFC-OL, and both exceed adaptive filters. But at high gains, DFC-IL clearly outperforms DFC-OL by maintaining system stability—a critical improvement for users with severe hearing loss who need high amplification.
For hearing aid manufacturers, this approach could enable smaller, more powerful devices without sacrificing feedback control. The technique also opens the door to training other audio systems that must handle unstable operating regimes.
- DFC-IL uses a two-stage training: pre-train on stable systems, then fine-tune on a wider gain range including unstable conditions.
- At high amplification gains, DFC-IL maintains stability while open-loop DFC and adaptive filters fail.
- Low-gain performance matches open-loop DFC and exceeds adaptive filters, with no trade-off in normal use.
Why It Matters
Enables hearing aids to deliver higher amplification without feedback, improving speech understanding for severe hearing loss.