Abstract: Aiming at the problem of low accuracy and poor robustness of speech recognition in complex noise environment, we proposed  a speech recognition method based on Mel cepstrum fusion feature of increasing and decreasing residuals.  This method first used the increase and decrease component method to screen the key speech features, and then mapped them to the Mel domain-residual domain spatial coordinate system to generate the increase and decrease residual Mel cepstral coefficients. Finally, these fusion features were used to train the end-to-end model. The experimental results show that the proposed method significantly improves the  accuracy and performance of speech recognition under different noise types and signal-to-noise ratio conditions. Under the low signal-to-noise ratio condition of -5 dB, the speech recognition accuracy reaches 73.13%, while the average speech 
recognition accuracy under other noise conditions reaches 88.67%, which fully proves the effectiveness and robustness of the proposed method.

Key words: speech recognition, residual Mel cepstral coefficient, feature screening, increase and decrease , component method