Robust Technologies towards Automatic Speech Recognition in Car Noise Environments

This paper presents the research on robust automatic speech recognition (ASR) in car noise environments. In the front-end design, speech enhancement technologies are used to suppress the background noise in frequency domain, and then spectrum smoothing is implemented both in time and frequency index to compensate those spectrum components distorted by noise over-reduction. In acoustic model training, we propose to use an immunity learning scheme, in which pre-recorded car noises are artificially added to clean training utterances with different signal-to-noise ratios (SNR) to imitate the in-car environments. After analyzing the SNR and noise spectrum of real in-car utterances, we further refine the immunity training set by adjusting the distribution of SNR and increasing the proportion of training noises that has a similar characteristic. Evaluation results of isolated phrase recognition show that the ASR system with proposed technologies achieves the average error rate reduction (ERR) of 90.68% and 79.08% for artificial car noisy speech and real in-car speech respectively, when compared with the baseline system in which no robust technology is used