Research Of Speech Recognition Based On Hybrid Model Of HMM And ANN

In recent years, we have made great progress in the non-specific continuous speech recognition technology, artificial intelligence and machine learning has become a very popular area of research. Research of speech recognition has made great progress in theory, but when the theory is applied to the actual product, there are two terrible problems. On the one hand, the influence of noise can’t be ignored. On the other hand, speech samples for training is usually limited, parameters of the model are not fully trained. Considering the two issues, we proposed two solutions. We introduce hybrid model composed by hidden Markov model and artificial neural networks and subspace Gaussian mixture model to solve these two problems. The main contents of the article includes the following:(1) This thesis introduces the basic principles of speech recognition technology, including pre-emphasis, framing windowing and endpoint detection. The article focuses on the endpoint detection, and we do a simulation experiment for the traditional dual threshold endpoint detection algorithm.(2) This thesis introduces the LPC, LPCC and MFCC in feature extraction of speech recognition. Finally, we do a comparative analysis for LPCC and MFCC, and choose MFCC factor as feature of speech recognition.(3) HMM speech recognition systems can only achieve better recognition results in pure speech environment. When speech recognition system is applied to the actual product, the speech training sample data is limited. This thesis introduces the subspace Gaussian mixture model, describe and analysis the principles of HMM+SGMM.(4) First, the thesis take a test for HMM and HMM+SGMM using different training sample. Secondly, taking an experiment for HMM+ANN in noisy environments. Finally, we apply improved endpoint detection to HMM+SGMM. The results show that: when voice training samples is limited, subspace Gaussian mixture model is an effective means of optimization model; HMM+SGMM model still has better recognition performance in noisy environment; improved endpoint detection is applicable for HMM+SGMM.(5) This thesis put forward a mixed model composed by hidden Markov and artificial neural network to apply to noisy environments, and take an experiment for HMM and HMM+ANN under SNR ranging from 5dB to 35 dB. The results show that the recognition rate of HMM+ANN model is higher than HMM. The thesis also take a test for improved endpoint detection algorithms, the results show that improved endpoint detection algorithms has a certain improvement for HMM and HMM+ANN.