In-vehicle Noise Cancellation Based on Probabilistic Recognition Switching Algorithm

The functionality of modern vehicles is becoming increasingly comprehensive, with electronic equipment configurations growing more complex. However, operating screens and other devices while the driver's hands are off the steering wheel undoubtedly poses safety risks. Installing precise human-machine interaction devices can address this issue. Nevertheless, the presence of various types of in-car noise diminishes the accuracy of speech recognition. Hence, speech enhancement technology in the automotive environment is crucial. To address the issue of various mid-to-low-frequency noises such as engine and ground friction affecting speech recognition and communication, this study has designed a combined architecture of bimodal probabilistic Wiener filtering and bimodal probabilistic spectral subtraction algorithms. The prior signal-to-noise ratio (SNR) serves as the switching criterion; when the prior SNR is less than 0.35, a Wiener filter is utilized, and when it exceeds 0.35, a subtraction filter is employed to minimize the "musical noise" left by spectral subtraction. The conditional probabilities in the switching framework are derived from Bayes' theorem to identify constructive and destructive noises. The final simulation results demonstrate that compared to various traditional algorithms, the proposed combined architecture achieves the optimal performance in removing in-car noise and enhancing speech quality.