Noisy Chinese digit string speech recognition based on tone modeling

It is attempted to utilize tone information to improve the performance of noisy Chinese digit string speech recognition. Multi-space probability distribution based HMM (MSD-HMM) is used to model the discontinuous tone features. The effect of noisy environment on tone features is analyzed and the feasibility of utilizing tone information to improve noisy speech recognition is discussed. Experimental results show that the proposed method can averagely obtain 17.2% relative reduction of digit error rate for the noisy data SNR from 5 dB to 20 dB, comparing with the method without tone information. The study concludes that it is effective to apply MSD-HMM based tone model to enhancing noisy Chinese digit string speech recognition.