A Chinese Speech Recognition System Based on Binary Neural Network and Pre-processing

Neural networks have made excellent progress in the field of speech recognition. However, more research needs to be done in some scenarios where computational resources are limited or real-time, and low power consumption is required. In this paper, we propose a lightweight speech recognition model based on pre-processing + binary neural network, which can significantly reduce the number of weight parameters while ensuring an acceptable error rate. The speech pre-processing part converts the 1D speech signal to the 2D Mel spectrum and uses Voice Activate Detection (VAD) to make the speech Mel spectrum input variable. The speech data set is also expanded using data augmentation methods. For convolutional layers, the weights are binarized to reduce the number of model parameters and improve computational and storage efficiency. The number of model parameters after quantization is 6.94% of the number of full precision model parameters, and the error rate on the ST CMD speech dataset increases by only 2.07%. Finally, a circuit structure based on binary weights for convolutional computation is designed, and a single multiplication can be implemented using only the hardware resources of the 7 Look Up Table (LUT).