Artificial intelligence is very popular in today's society.With the close connection between speech recognition technology and artificial intelligence,the research heat is on the rise.The advantages of speech recognition technology,such as convenient use,high user acceptance and low cost of sound input equipment,make it widely used.However,the traditional speech recognition algorithm has poor anti-noise ability and low recognition accuracy,so it is not competent for practical application.In order to comply with the trend of modernization and ensure the accuracy of speech recognition,this paper first proposed a double sparse K-SVD algorithm based on the K-SVD algorithm with non-negative constraints,this algorithm has better convergence,and its relative error is smaller than K-SVD algorithm.The use of the double sparse K-SVD algorithm for sparse representation of the dictionary can allow the dictionary to obtain more ideal adaptive ability,improve the calculation speed,and contribute to the sparse representation of larger signals.Although the traditional joint dictionary can obtain the signal distribution characteristics and rules in the pure speech and noise samples,it ignores the existence of correlation atoms between the pure speech dictionary and the noise dictionary,which leads to the source confusion in the sparse reconstruction stage.Therefore,a speech denoising algorithm based on discriminative joint dictionary learning is proposed in this paper.By adding dictionary discriminating constraint items among dictionaries,this algorithm enhances the discriminating ability of dictionaries,enables each signal to be sparsely represented in the corresponding sub-dictionary of the joint dictionary and inhibits its sparse representation in the non-corresponding sub-dictionary.At the same time,the double sparse K-SVD algorithm was applied to avoid the existence of correlation atoms between dictionaries and improve the adaptability of the algorithm.The experimental analysis proves the effectiveness of the discriminative joint dictionary,and the algorithm still has a strong ability to reduce noise under the interference of high noise.Finally,the hardware of embedded speech recognition system is designed for practical application scenes.The combination of traditional LD3320 speech recognition chip and 51 microcontroller limits the number of recognition instructions and the recognition success rate is not high,so the V290 pub speech recognition module combined with STM32 microcontroller is selected to replace it,The feasibility of the hardware design is verified by the recognition success rate experiment,and the combination of the algorithm proposed in this paper and the embedded platform design has a good voice noise reduction ability through the oscilloscope waveform analysis of the speech signal. |