Speech recognition method based on ADRMFCC fusion features

The invention provides a speech recognition method based on ADRMFCC fusion features in order to solve the problems of low speech recognition accuracy and poor robustness in a complex noise environment. According to the method, speech contribution degrees of all dimension features of a residual Mel-frequency cepstral coefficient (RMFCC) and a Mel-frequency cepstral coefficient (MFCC) are screened by using a component increase and decrease method to improve speech recognition performance, then the screened features are spliced and fused, and finally the processed fused feature ADRMFCC is sent to a bidirectional recurrent neural network for recognition. Experimental results show that the recognition accuracy and performance of the method provided by the invention are far higher than those of a single feature under the conditions of different noise types and signal-to-noise ratios, the recognition accuracy can reach 73% or above under the condition of a low signal-to-noise ratio of-5dB, and the average accuracy u