Research on Speech Recognition Acceleration Technology Based on Embedded Platform

This paper implements the latest speech recognition technique and improves it so that it can be used on embedded mobile devices. In order to implement the speech recognition system, I build the two-layer Long Short-Term Memory networks (LSTMs) using the TensorFlow framework and train the model by using the Mel Frequency Cepstral Coefficients (MFCCs) vector extracted from the wave files. In order to speeds up the model, this paper innovatively applies the pruning method to the LSTMs by pruning the weights of each cell and regularizing the weight matrix of regression layer. After pruning, I retrain the model to preserve the accuracy. Finally, this new model discussed in this paper successfully improves the performance of the original model without reducing the precision of the model. It reduces the computational complexity and speeds up the speech recognition process.