Speech Enhancement Under High-speed Driving Environment

ABSTRACT:Speech enhancement technology used to eliminate noise in speech signal and improve signal quality is widely applied in industrial production and our daily life Especially in the automotive industry, speech enhancement is one of the key technologies in the design of the car voice control system. Many existing speech enhancement methods have a good effect on the car which runs at low-speed, on the contrary, they show poor performance if the car runs at high-speed.In the high-speed driving conditions, the car noise is mainly constituted by the wind noise and tire noise, and the noise characteristics are substantially similar when the cars run on the same road at the same speed. Therefore, we proposed a noise estimation based on the acoustic modeling.Firstly, we propose a noise estimation method utilizing the noisy speech power spectrum to establish the probability of Gaussian mixture model(GMM). We choose the optimal model size by comparing the training time and estimation accuracy of the GMM models in defferent sizes, the parameters of which are calculated by the expectation maximization(EM) algorithm. Then we predict the noise power spectrum according to the minimum mean square error(MMSE) criterion to lay the groundwork for the following speech enhancement. Secondly, we utilize an improved algorithm for computing a priori SNR and modify the relevant parameter. We rewrite the formula by combing two calculations of a priori SNR with the above noise estimation, and complete the update of the clean speech power spectrum.Simulation results show that the proposed algorithm could remove the car noise in the high speed driving environment effectively. In respect of noise estimation, the average error of proposed algorithm is20%less than that of conventional algorithms, while in other respect of speech enhancement, the signal-to-noise ratio(SNR) of proposed algorithm is at least1dB more than that of conventional algorithms.