Acoustic feedback cancellation in speech reinforcement systems for vehicles

Alfonso Ortega, Eduardo Lleida, Enrique Masgrau, Luis Buera, Antonio MiguelInterspeech 2005 · DOI

Abstract

Speech reinforcement systems in vehicles often suffer from acoustic feedback due to the closed-loop nature of the microphone-speaker setup. Standard Acoustic Echo Cancellation (AEC) techniques based on the Minimum Mean Square Error (MMSE) criterion fail in these environments because the input signal is correlated with the error signal, leading to biased filter estimates. This paper proposes a decorrelation approach using adaptive whitening filters within a Filtered-X Least Mean Squares (FX-LMS) framework. Experimental results using real car acoustic paths demonstrate that the proposed method significantly reduces identification error and improves system stability compared to conventional AEC.

Introduction

In-car communication systems are designed to help passengers hear each other over ambient road and engine noise. However, the proximity of microphones and speakers within the small cabin creates a closed-loop system. This leads to the "howling" effect, where the speaker output is picked up by the microphone and re-amplified. While standard AEC is effective for open-loop echo cancellation (like in telephony), it fails in closed-loop reinforcement systems. The mathematical "incestuousness"—where the filter's input is a delayed version of its own error—biases the adaptive filter, preventing accurate identification of the acoustic path.

Methodology

The authors analyze the failure of the MMSE criterion in closed loops, identifying that error is driven by system gain ($K$), propagation delay ($\Delta$), and signal spectral characteristics. To address this, they introduce a Filtered Error approach.

Proposed system architecture incorporating whitening filters $p(n)$ and the FX-LMS algorithm.

The core innovation is the use of whitening filters $p(n)$ to decorrelate the signals. By "whitening" the error signal, the predictable components of speech are removed, allowing the adaptive filter to focus on the unpredictable "innovation." This transforms the system into an FX-LMS algorithm. Since speech and noise are non-stationary, these whitening filters are updated adaptively using linear prediction.

Experimental Setup

The study utilized a mix of simulated and real-world data to validate the approach:

  • Real-World LEM Path: A 600-coefficient impulse response measured inside a medium-sized car.
  • Speech & Noise: Real sentences sampled at 8 kHz, combined with actual car noise recorded while driving on a highway, achieving a signal-to-noise ratio of approximately 10 dB.

Results and Discussion

A key finding is that perfect path identification in a closed loop only occurs naturally if the excitation is white noise. For structured signals like speech, decorrelation is mandatory. The researchers measured performance using the normalized $l_2$ norm of the weight misadjustment, which quantifies the distance between the filter's estimate and the true acoustic path.

Performance comparison of weight misadjustment (||ε||²) with and without the whitening filter across varying gain factors K.

As shown in Figure 6, the proposed whitening filter consistently achieves a lower estimation error compared to standard filters across all tested gain levels. Furthermore, the study found that increasing loop delay ($\Delta$) helps reduce identification error by naturally decreasing the correlation between the current error and the filter input. Higher amplification ($K$) generally improves the "echo-to-excitation" ratio, which can help the filter converge more accurately, provided the system remains stable.

Conclusion

The integration of adaptive whitening filters effectively breaks the feedback loop that biases standard AEC in vehicle speech reinforcement. While this study focused on single-channel systems, future work should address multi-channel configurations and the computational demands of running multiple adaptive filters on automotive hardware. Additionally, addressing non-linearities introduced by speaker distortion at high volumes remains an area for further research.

Summary updated

BibTeX
@inproceedings{ortega2005acoustic,
  author = {Alfonso Ortega and Eduardo Lleida and Enrique Masgrau and Luis Buera and Antonio Miguel},
  title = {{Acoustic feedback cancellation in speech reinforcement systems for vehicles}},
  booktitle = {Interspeech},
  year = {2005},
  doi = {10.21437/interspeech.2005-675},
}
Figure