Research And Application Of A Fast Isolated Word Recognition Method Based On Embedded

In order to make machinery understand the voice, we need to do speech recognition at the firstly. Hence, the speech recognition is an important branch in intelligent technology. With the development of HMM,the speech recognition technology comes into people’s lives gradually. However, the speed of development of subsequent theoretical is relatively slow and the price of speech recognition chip is much expensive, which affected the widely use of speech recognition. The inaccuracy of the extraction of speech features and the complication of the speech recognition model are the two main problems that we have to face.In order to deal with this issue, an improved algorithm has been proposed in this paper.In order to study the parameters that affect the speech recognition, we did the related feature extraction experiments. We found that the key of the speech recognition feature is the time-frequency domain feature by analyzing of the existing time-domain characteristics, frequency domain features and time- frequency domain features. By simplifying the spectrogram, we get a new time-frequency characteristics – zero-spectrum. A large number of experiments show that the zero-crossing spectroscopy is a simple characteristic function which has high recognition efficiency.The existing speech recognition model has high recognition precision but its computation is much complicated. In order to solve this problem, we proposed a new speech recognition model with high recognition precision and low computation complexity called biomimetic pattern recognition algorithm based on the hyper-sphere string. Compared with the Dynamic Time Warping(DTW) and Hidden Markov Model(HMM) via repeated experiments, we can know that the algorithm complexity of biomimetic pattern recognition is much lower than other algorithms while the accuracy of recognition of the proposed model is not lower than other algorithms at the same time.As for the existing speech recognition models, the process of extraction of the characteristic parameters is complex and is difficult to compute. At the same time, the computing process of the existing models is time consuming, which makes it difficult to apply on low performance embedded system. Furthermore, the problem lead to the high price of voice chip and speech recognition system. Aiming at the problems above, a new method of feature extraction was proposed in this paper and biomimetic pattern recognition was used to recognize at the same time. By do the simulation experiment using the MATLAB software, we obtained the result that the proposed method has obvious advantages of low algorithm complexity, fast-running and high recognition rate. Finally, we implemented the isolated word recognition by transplanting the algorithm to the STM32 hardware.This study of this paper has extensive application value. With the increasingly demanding of people’s quality of life, it acquires comprehensive advantage in low cost, execution rate, and so on. Hence, the work we did in this area is a beneficial attempt.