Method and device for speech recognition by use of LSTM recurrent neural network model

Abstract

The invention discloses a method and a device for speech recognition by use of a long-short term memory (LSTM) recurrent neural network model. The method comprises the following steps: receiving speech input data at the (t)th time; selecting the LSTM hidden layer states from the (t-1)th time to the (t-n)th time, wherein n is a positive integer; and generating an LSTM result of the (t)th time according to the selected at least one LSTM hidden layer state, the input data at the (t)th time and an LSTM recurrent neural network model. With the method and the device, the 'tail effect' of the deep neural network is well solved, and the accuracy of speech recognition is improved.

Description

technical field [0001] The present invention relates to the technical field of speech recognition, in particular to a method and device for performing speech recognition using an LSTM (Long-Short Term Memory, long-short-term memory) cyclic neural network model. Background technique [0002] Speech recognition technology refers to the process of converting an input speech signal into a text output, usually including an acoustic model, a language model, and a corresponding decoding search method, and its performance largely depends on the construction of an acoustic model. Since speech is a typical temporal signal, recurrent neural networks, especially long short-term memory LSTM recurrent neural networks, have gradually become a new direction of acoustic modeling in speech recognition because of their strong temporal modeling capabilities. [0003] However, the speech signal has short-term stationary characteristics, and the difference between adjacent frames obtained by feat...

AI Technical Summary

Problems solved by technology

[0003] However, the speech signal has short-term stationary characteristics, and the difference between adjacent frames obtained by feature extraction (for example, window length 25ms, step size 10ms) is often small. Repeated occurrences will cause the "after-tail effect", that is, frames with simple patterns (such as silent frames) are continuously input into the network for dozens of frames, which will cause the recurrent neural network to have a strong memory of the simple pattern, while frames with different labels When inputting, it still cannot be adjusted quickly, resulting in estimation errors and affecting recognition performance, for example, if figure 1 As shown, because the pattern corresponding to label 1 is relatively simple but repeats a lot of time beats, when the new feature of label 3 is input into the network, the network cannot respond for a long time, so the subsequent three frames are all wrongly predicted into tab 1

Images