TL;DR
This paper introduces a novel speech enhancement system, the BC-assisted Generalized Sidelobe Canceller (BCA-GSC), which leverages a bone-conduction (BC) sensor alongside traditional microphones. By using the BC sensor for highly accurate voice activity detection (VAD) even in noisy environments, the system significantly improves the performance of the adaptive blocking matrix (ABM) and adaptive noise canceller (ANC) components of the GSC, leading to superior speech quality and intelligibility.
What Problem Does This Paper Address?
Speech enhancement systems, particularly those using microphone arrays like the Generalized Sidelobe Canceller (GSC), aim to isolate and improve the quality of desired speech signals while suppressing unwanted noise. Conventional GSCs face challenges in accurately distinguishing between speech and noise, especially in highly non-stationary noise conditions or low signal-to-noise ratio (SNR) environments. This can lead to the leakage of desired speech into the noise estimate or the cancellation of speech itself. The paper addresses the limitations of existing GSC methods by proposing a more robust approach that utilizes the unique properties of a bone-conduction (BC) sensor to enhance the critical voice activity detection (VAD) process, which in turn improves the control of the GSC's adaptive components.
Methods
The core innovation is the BC-assisted GSC (BCA-GSC) algorithm. It builds upon the conventional GSC structure, which consists of a Fixed Beamformer (FBF), an Adaptive Blocking Matrix (ABM), and an Adaptive Noise Canceller (ANC). The key enhancements are:
- Bone-Conduction (BC) Sensor for VAD: The BC sensor, being less susceptible to ambient acoustic noise than air-conduction (AC) microphones, is used to derive a highly accurate Voice Activity Detection (VAD) signal. This VAD signal is crucial for controlling the adaptive components.
- VAD-Controlled ABM: The VAD information from the BC sensor is used to control the adaptation of the ABM. This ensures that the ABM is more effectively trained to block desired speech and pass noise, reducing signal leakage. The paper employs a Recursive Least Squares (RLS) algorithm for updating the ABM coefficients, which offers efficient convergence.
- VAD and SIR-Controlled ANC: The ANC's adaptation is jointly controlled by the VAD information and the Signal-to-Interference Ratio (SIR). This dual control mechanism allows for more precise noise suppression, particularly by adapting the ANC during speech absence and adapting the ABM during speech presence.
- Iterative Refinement: The output of the ANC is fed back as a reference signal to the ABM for a single iteration. This step further refines the noise suppression by allowing the ABM to adapt to the residual noise characteristics after ANC processing.
Datasets
The paper does not explicitly mention specific publicly available datasets for training. Instead, it describes the experimental setup where noisy AC speech signals were generated by corrupting clean AC speech with various types of noise (directional and diffuse) at different SNRs (0, 5, and 15 dB). Clean AC speech and BC speech signals were recorded simultaneously in a silent environment.
Key Insights and Ideas
- BC Sensor as a "Clean" VAD Source: The most significant insight is the effective use of a BC sensor as a reliable source for VAD, even in extremely noisy conditions where AC microphones struggle. This bypasses a major bottleneck in traditional GSC systems.
- VAD as a Master Switch for Adaptive Components: The paper demonstrates that precise VAD control is paramount for optimizing the performance of both the ABM and ANC. By using VAD to dictate when each component should adapt, the system achieves better noise suppression and signal preservation.
- Targeted Adaptation: The strategy of adapting the ABM during speech presence and the ANC during speech absence is a clever way to leverage the distinct roles of these components, leading to improved overall performance.
- Limited Benefit of Multiple Iterations: The finding that a single iteration of feeding ANC output back to the ABM provides significant improvement, but further iterations yield diminishing returns, is a practical insight for system design.
Key Results
The proposed BCA-GSC algorithm demonstrated significant improvements in speech enhancement compared to the conventional GSC:
- Objective Quality: The paper reports improvements in Perceptual Evaluation of Speech Quality (PESQ) scores across various noise environments and SNRs. For instance, in the case of 90° music noise at 5 dB SNR, BCA-GSC 1 (no iteration) achieved a PESQ score of 95.43, compared to 92.86 for GSC 1. BCA-GSC 2 (one iteration) further improved this to 95.80.
- Intelligibility: Short-Time Objective Intelligibility (STOI) scores also showed enhancement. For 90° music noise at 5 dB SNR, BCA-GSC 1 achieved 95.43% STOI, and BCA-GSC 2 reached 95.80%, significantly outperforming GSC 1's 92.86%.
- Noise Suppression: Visual inspection of spectrograms (Figure 6) showed that the BCA-GSC effectively suppressed noise, particularly in regions where the conventional GSC left residual noise or even cancelled desired signal components.
- Robustness: The algorithm showed robustness in various noise conditions, including directional and diffuse noises, and across different SNR levels.
Limitations and Future Work
- Computational Complexity: While the paper mentions using RLS for efficient convergence, it doesn't provide a detailed analysis of the computational complexity of the entire BCA-GSC system compared to the conventional GSC. This could be a consideration for real-time embedded applications.
- BC Sensor Integration: The reliance on a BC sensor adds hardware complexity and cost compared to purely microphone-based systems. The paper assumes the BC sensor is already embedded, as in wearable devices.
- Future Work: The authors suggest that further research could explore more advanced adaptive filtering algorithms or investigate the impact of different BC sensor placements and characteristics on performance. The potential for using the BC sensor for other aspects of speech processing, beyond VAD, could also be an avenue for future exploration.
Summary updated