In this chapter, we describe a method of minimum mean-squared error (MMSE) a posteriori estimation of high-variance vehicular noise. The proposed method considers spectral instances of noise as sampled values from a stochastic noise process and estimates them with given statistical properties of noise and current noisy observation. Accuracy of the noise estimation method is evaluated in terms of the accuracy of a spectrum-based voice activity detection (VAD), in which speech presence is determined by the a priori and a posteriori signal-to-noise ratios (SNRs) in each frequency bin. VAD experiments are performed on clean speech data by adding four different types of vehicular noise, each with the SNR varying from -10 to 20 dB. Also, isolated digit recognition experiments are performed using original noisy recordings from the AVICAR corpus. Experimental results show that the proposed noise estimation method outperforms both the MMSE a priori noise estimation and autoregressive noise adaptation methods especially for low SNR.