Speech recognition method based on simplified and improved Transform model

Abstract

The invention relates to a speech recognition method based on a simplified and improved Transform model, and belongs to the technical field of speech recognition. The method comprises the following steps: A, preprocessing a voice signal, and extracting a multi-dimensional mfcc Fbank; B, carrying out convolution on the extracted Fbank features by using a CNN (Convolutional Neural Network); C, inputting the features obtained after convolution into the improved Transform network structure; D, taking the CTC loss as a loss function of the acoustic model, performing prediction by adopting a Beam search algorithm, and performing optimization by using an Adam optimizer; E, starting to train the model to continuously iteratively optimize the model to an optimal model structure; and F, carrying out model verification. According to the method, the parameter quantity of the model is greatly reduced on the premise of ensuring that the word error rate is reduced, the calculation speed and the convergence speed of the original model are improved, the training time of the model is shortened, and the working efficiency of speech recognition is effectively improved.

Description

technical field [0001] The invention relates to a voice recognition method based on a simplified and improved Transformer model, belonging to the technical field of voice recognition. Background technique [0002] Transformer, as a new deep learning algorithm framework, has attracted the attention of more and more researchers and has become a current research hotspot. The self-attention mechanism in the Transformer model is inspired by humans who only focus on important things, and only learn important information in the input sequence. For the speech recognition task, the focus is on transcribing the information of the input speech sequence into the corresponding speech text. In the past, the speech recognition system was composed of acoustic model, pronunciation dictionary and speech model to realize speech recognition tasks, while Transformer can integrate acoustics, pronunciation and speech models into a single neural network to form an end-to-end speech recognition sys...

AI Technical Summary

Problems solved by technology

[0006] The Transformer model with a deep network structure is more likely to fall into a local optimum, making it more difficult to train the Transformer-based end-to-end speech recognition model

Secondly, due to the self-attention mechanism in Transformer, it has high complexity when computing in parallel between frames

Therefore, the Transformer with a deep structure has a large number of parameters and requires powerful computing power, and there is a problem that the model cannot be run in the deployment of edge devices and embedded devices.

Images