A Noise Robust Front-End with Low Computational Cost for Embedded In-Car Speech Recognition

This paper proposes a noise robust front-end with low computational cost for embedded in-car speech recognition. The minimum mean-square error (MMSE) estimation algorithm is adopted to suppress the background noise, and in the gain function calculation a suitable piece-wise linear function is used to substitute the traditional Taylor series accumulation method to simplify the computation complexity. After speech enhancement, spectrum smoothing is implemented in both time and frequency index with geometric sequence weights to further compensate the spectral components distorted by noise over-reduction. Experiments on Chinese isolated phrase recognition show that the proposed front-end significantly improves the recognition robustness in car environments while the computational load is extremely reduced. Compared with the ETSI advanced front-end, the average error reduction rate (ERR) of 12.2% and 4.5% is obtained in artificial car noisy speech and real in-car speech, respectively.