In vehicular networks, V2V/V2I communications are highly susceptible to wireless channel noise, which compromises network security. Wavelet threshold denoising can effectively remove environmental noise before extracting channel features, thereby mitigating interference with physical-layer encryption and improving signal recognition accuracy. This paper proposes a hybrid speech recognition method combining wavelet threshold denoising with ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise). The wavelet threshold denoising processes high-frequency modal components, while ICEEMDAN handles low-frequency modalities. Furthermore, we introduce an adaptive thresholding mechanism and an improved threshold function that simultaneously ensures continuity at threshold points and addresses wavelet coefficient bias.