Abstract
This paper proposes a speech enhancement algorithm featuring a novel interaction between adaptive beamforming and multi-channel post-filtering. A subband feedback controller, driven by speech presence probability (SPP), is integrated into a Generalized Sidelobe Canceller (GSC) framework. This design enhances robustness in adverse environments and mitigates the common issue of signal cancellation. Multi-channel post-filtering is employed to suppress diffuse noise and transient interferences while providing critical SPP information to the beamformer. Experimental results demonstrate significant improvements in signal preservation and noise reduction compared to existing algorithms.
Introduction
Microphone arrays typically utilize the Generalized Sidelobe Canceller (GSC) to isolate a target speaker while suppressing directional interference. However, GSC systems often suffer from "signal cancellation," where misalignment or reverberation causes the adaptive filter to erroneously suppress the desired speech. Additionally, while beamformers excel at directional noise suppression, they are less effective against diffuse noise, such as ambient room hum. Traditionally, beamformers and post-filters operate independently. This research introduces a closed-loop system where these components communicate to optimize performance.
Proposed Methodology
The proposed architecture integrates three key innovations to address the limitations of standard GSC systems.
1. Enhanced Directional Suppression
The system employs a modified Blocking Matrix (BM) that utilizes multiple sensor pairs rather than just adjacent microphones. This approach captures richer spatial information, resulting in a more accurate "noise-only" reference channel, which is vital for preventing the GSC from affecting the target signal.
2. Auditory Subbanding
Signals are processed in subbands grouped according to the Bark scale. By mimicking human auditory perception, the system allocates higher resolution to low-frequency bands where speech energy is concentrated, improving both computational efficiency and processing quality.
3. The SPP Feedback Loop
The core innovation is the interaction between the multi-channel post-filter and the adaptive beamformer. The post-filter calculates the Speech Presence Probability (SPP) for each subband. This information is fed back to a Subband Feedback Controller (SFC) that dynamically adjusts the learning rate (step-size) of the adaptive filter. When the SPP is high, the learning rate is reduced to prevent the filter from adapting to—and thus canceling—the desired speech.
Experimental Results
The algorithm was evaluated using 20 sentences from the TIMIT database in a simulated reverberant room ($T_{60} = 300$ms). Noise conditions included a competing speaker at 90° and white Gaussian noise at 45°, with global SNRs ranging from -6dB to 6dB.
The proposed method was compared against standard time-domain GSC (GSC-TD), frequency-domain GSC (GSC-FD), and Cohen’s post-filtering method. Performance was measured using Perceptual Evaluation of Speech Quality (PESQ-MOS), Noise Reduction (NR), and Log-Spectral Distance (LSD).
As shown in the table above, the proposed method consistently achieves the highest PESQ-MOS scores, reaching 2.98 at 0dB SNR compared to 2.13 for GSC-TD and 2.85 for Cohen’s method. Furthermore, the system maintained a high noise reduction of approximately 35dB across varying SNR levels and exhibited the lowest spectral distortion (LSD), indicating superior naturalness in the processed speech.
Conclusion
By establishing a feedback mechanism between the post-filter and the beamformer, this research successfully addresses the signal cancellation problem in GSC systems. The use of auditory subbanding and an improved blocking matrix further enhances robustness against both directional and diffuse noise. Future work may explore the application of this feedback control to diverse microphone array geometries and higher sampling rates for modern communication standards.
Summary updated