End-to-end speech recognition method based on adaptive learning rate

Abstract

The invention provides an end-to-end speech recognition method based on an adaptive learning rate. The method particularly comprises steps: (1) a bidirectional recurrent neural network is adopted as an acoustic model, a forward recurrent neural network hidden layer h-right arrow and a reverse recurrent neural network hidden layer h-left arrow are calculated respectively, a long-short term memory cell (LSTM) is then adopted to replace the h-right arrow and the h-left arrow, and final output, that is y, of the neural network is obtained; (2) an acoustic model modeling unit in the first step serves as a phoneme, a connectionist temporal classification criterion is adopted, a blank symbol is introduced to assist alignment, an objective function is built and calculated, partial derivative computation is carried out on the objective function relative to the neural network output, an error back propagation algorithm (BP) is then used to calculate the gradient g of a parameter set w in a weight matrix in the first step; and (3) based on the first-order gradient information, that is, the gradient g provided by the second step and in combination of an ADADELTA adaptive learning rate method, the parameter set w is updated.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to an end-to-end speech recognition method based on an adaptive learning rate. Background technique [0002] With the rise of deep learning, speech recognition technology based on deep neural networks has made remarkable progress. At present, the commonly used speech recognition method adopts a hybrid method based on hidden Markov model and deep neural network. It needs to train hidden Markov model and corresponding Gaussian mixture model to provide frame-level training annotations for subsequent training of deep neural network. However, the speech recognition framework based on the hybrid method of hidden Markov model and deep neural network is more complicated: first, the Gaussian mixture model trained by it will not be used in the final decoding process; second, the training model relies on too much Linguistic knowledge, such as the problem set needed to build a decis...

AI Technical Summary

Problems solved by technology

[0004] The object of the present invention is to provide a kind of end-to-end speech recognition method based on adaptive learning rate in or

Images