Lightweight end-to-end speech recognition method based on convolutional self-attention transformation network

Abstract

The invention discloses a lightweight end-to-end speech recognition method based on a convolutional self-attention transformation network, and the method comprises the steps: firstly constructing a lightweight end-to-end speech recognition model based on the convolutional self-attention transformation network, improving the convolutional self-attention transformation network through the model, and forming an efficient convolutional self-attention transformation network; applying the low-rank decomposition to a feed-forward layer in the convolutional self-attention transformation network to form a low-rank feed-forward module; providing a multi-head efficient self-attention MHESA, and adopting the MHESA to replace dot product self-attention in a convolutional self-attention transformation network encoder; and finally, obtaining a voice recognition model through training to recognize the voice. According to the method, the calculation complexity of the self-attention layer of the encoder is reduced to be linear, the parameter quantity of the whole model is reduced by about 50%, and the performance is basically unchanged.

Description

Technical field [0001] The present invention belongs to the field of pattern recognition, and more particularly to a lightweight end-to-end speech recognition method. Background technique [0002] Speech Identification (ASR, Automatic Speech Recognition) aims to convert voice signals into text content, which can image the "machine auditory system", is an important research area of ​​human-computer communication and interaction technology, and is also a key technology of artificial intelligence. one. Voice recognition can be applied to many aspects including voice assistants, automatic driving, smart home, handheld mobile devices. In recent years, end-to-end voice recognition technology has many advantages compared to traditional methods. If the labeling of training data is simple, the dependence of linguistics is small, and there is no need to hidden Markov chain in the hidden Markov chain. The conditions of the transfer probability are independent assumptions, and the training a...

AI Technical Summary

Problems solved by technology

However, Transformer also has some shortcomings, such as the computational complexity of dot product self-attention increases quadratically with the length of the input feature sequence, and the number of model parameters is large, etc.

Images