TL;DR
This paper introduces a novel, efficient method for detecting and suppressing "howling" (acoustic feedback) in digital hearing aids. By employing a segmented notch filtering approach with a closed-form FIR filter design, the method achieves high attenuation of howling while maintaining linear phase characteristics and low computational complexity, making it highly suitable for real-time hearing aid applications.
What Problem Does This Paper Address?
Hearing loss is a growing concern, and hearing aids are a primary solution. However, a common and highly disruptive issue in digital hearing aids is "howling," an unpleasant, high-intensity feedback sound that occurs when the amplified sound from the speaker re-enters the microphone, creating a positive feedback loop. Existing methods for suppressing this howling, such as gain attenuation and adaptive echo cancellation, often suffer from drawbacks like insufficient attenuation, slow response times, nonlinear distortion, or high computational cost. This paper aims to address these limitations by proposing a more effective and efficient howling suppression technique.
Methods
The proposed scheme is based on a segmented notch filtering approach. The core innovation lies in a closed-form FIR notch filter design that meets several critical requirements: low complexity, arbitrary notching frequency control, high attenuation at the notch frequency, and linear-phase characteristics.
Here's a breakdown of the key components:
- Segmentation and Splicing: The incoming audio signal is divided into segments. Each segment is processed independently by the howling detection and suppression module. After processing, the segments are spliced back together to form the output signal.
- Energy Detector: A simple and computationally inexpensive detector that monitors the energy of each segment. If the energy exceeds a predefined threshold, it signals the potential presence of howling.
- Notch Frequency Estimator: When howling is detected, this module estimates the howling frequency. It uses a ratio-based hamming-windowed spectrum corrector, which involves computing the DFT, identifying the peak spectral position, and calculating an amplitude ratio to refine the frequency estimate.
- Closed-Form FIR Notch Filter: This is the heart of the suppression system. The filter coefficients are derived using a closed-form formula, which allows for efficient calculation and ensures linear-phase characteristics. The design incorporates a "reversed phase-shift combination" to achieve arbitrary notching frequencies.
- Data Extension: To mitigate nonlinear-phase effects introduced by the segmentation and FIR filtering process, extra data is extended at the beginning and end of each segment. This ensures that the output signal maintains its linear-phase characteristic.
Datasets
Not discussed in the paper. The experiments appear to use synthesized audio signals with added howling components.
Key Insights and Ideas
- Closed-form FIR filter for arbitrary notch frequencies: The ability to design a notch filter with coefficients derived from a simple analytical formula, while allowing for precise control over the notch frequency, is a significant advancement. This bypasses the complexity of iterative design methods.
- Linear-phase characteristic is achievable: By combining the closed-form FIR filter design with segmentation, splicing, and data extension techniques, the authors demonstrate that linear-phase distortion, a common issue in feedback cancellation, can be effectively avoided.
- Robustness to howling frequency: The proposed method shows consistent performance (measured by RSNR) across a range of howling frequencies, indicating its insensitivity to the exact frequency of the feedback.
- Low complexity is key for hearing aids: The paper emphasizes that the proposed scheme's low computational complexity makes it highly practical for implementation in real-time hearing aid systems, which have strict power and processing constraints.
Key Results
- The proposed notch filter achieves an attenuation value of over -330 dB at the howling frequency, significantly outperforming conventional IIR notch filters (around -150 dB).
- The recovered signal-to-noise ratio (RSNR) is consistently around 22 dB across various howling frequencies, filter orders, and data extension lengths, demonstrating the robustness and effectiveness of the scheme.
- Experimental results show that the howling component is completely removed in the recovered waveform, with no visible residual howling.
- The method successfully avoids acute jumps at segment edges, confirming its linear-phase characteristic.
Limitations and Future Work
- Limitations:
- The paper focuses on suppressing the howling effect itself, acknowledging that other distortions arising from the feedback channel might still be present, though their amplitudes are typically much smaller.
- While the method is robust to howling frequency, the paper doesn't extensively explore its performance under highly dynamic or rapidly changing acoustic environments beyond the described howling scenarios.
- Future Work:
- The authors suggest that the proposed scheme has vast potential for practical applications in future digital hearing aids, implying further development and integration into actual devices.
- The high flexibility of functionality updating and low power consumption suggest suitability for system-on-chip (SOC) implementation.
Summary updated