Speech recognition method, device, computer equipment and computer-readable storage medium

Abstract

The present invention provides a voice recognition method, device, computer equipment, and computer-readable storage medium, comprising: performing down-sampling processing on acquired audio data to obtain audio down-sampling data; dividing the audio down-sampling data into training audio data and Test audio data; sparsely process the weights in the convolutional layer of the binary convolutional neural network and the fully connected layer to obtain a sparse binary convolutional neural network; use the training audio data to perform sparse processing on the sparse The binarized convolutional neural network is trained to obtain the trained binarized convolutional neural network; using the test audio data, speech recognition is performed based on the trained binarized convolutional neural network. Since this scheme sparsely processes the weights in the convolutional layer and the fully connected layer of the binary convolutional neural network, it can save a lot of computing space and time.

Description

technical field [0001] The present invention relates to the technical field of voice recognition, in particular to a voice recognition method, device, computer equipment and computer-readable storage medium. Background technique [0002] Speech recognition has officially entered all aspects of our daily life and will become a new way for humans to interact with computers. For example, mobile phones, game consoles, and smart homes all require speech recognition. Speech recognition has been developed for decades and suddenly becomes hot. , which is attributed to deep learning. However, with the continuous improvement of the prediction accuracy of the neural network, the required storage space and the amount of calculation are also increasing, and the demand for hardware resources is also increasing. The large storage space and large amount of calculation required by the neural network seriously hinder its application in For applications on devices such as mobile phones, watch...

AI Technical Summary

Problems solved by technology

But even such a neural network has a little resistance to accurate, fast, low-latency, small model and low-power speech recognition, because the binarization method can know that there is no sparsity in the parameter part, and the parameter is not sparse. save space and time

Images