An efficient robust automatic speech recognition system based on the combination of speech enhancement and log-add HMM adaptation

Abstract

The acoustic mismatch between testing and training conditions is known to degrade severely the performance of automatic speech recognition (ASR) systems. The development of noise robust speech recognition algorithms is becoming increasingly important as speech technology is currently widely applied in real world applications. This paper presents a new efficient robust ASR system, which combines speech enhancement with log-add (LA) model adaptation. In the front-end stage, speech enhancement is adopted to suppress the additive noise imposed on the speech signal. Then, an LA model adaptation method is exploited to adjust the mean parameters of the hidden Markov models (HMM) to deal with the residual noise after speech enhancement processing. Experimental evaluations show that the proposed robust ASR system can achieve significant improvement in recognition across a wide range of signal-to-noise ratios (SNR), especially in very noisy environments.