Acoustic modeling method based on gated recurrent unit

Abstract

The invention discloses an acoustic modeling method based on a gated recurrent unit. The acoustic modeling method comprises the following steps: step 1, extracting corresponding acoustic features fromoriginal audio data; step 2, improving the gated recurrent unit by utilizing layer normalization, and calculating forward output of a neural network by utilizing the improved gated recurrent unit; step 3, training a model according to the state vector, calculated in the step 2, of the current moment; and step 4, decoding the trained model to find an output sequence with the maximum probability. According to the invention, the layer normalization technology is applied to the gated recurrent neural unit, so that the activation value of neurons can be normalized, the network convergence speed isimproved, and the network training time is reduced; an activation function in a traditional gated recurrent unit is replaced with an ELU activation function, so that the data robustness is improved;meanwhile, by optimizing a computational formula of the gate structure, model parameters of a traditional gated recurrent unit are reduced, and the recognition performance of the model can be improved.

Description

technical field [0001] The invention belongs to the technical field of speech recognition and relates to an acoustic modeling method, in particular to an acoustic modeling method based on a gated recurrent unit. Background technique [0002] In recent years, with the continuous development of artificial intelligence and computer technology, deep learning technology has been widely used in image, voice and other fields. As one of the most natural interaction interfaces between machines and humans, speech has become a research hotspot in academia and industry. [0003] The acoustic model is one of the core modules of the speech recognition system, and its performance directly affects the entire speech system. Before 2009, the basic structure of the acoustic model was a Gaussian Mixture Model-Hidden Markov Model (GMM-HMM). However, with the successful application of neural networks in the field of speech recognition, the traditional GMM- HMM is gradually replaced by DNN-HMM (...

AI Technical Summary

Problems solved by technology

[0005] However, in practical applications, this type of method is still far from the requirements of large-scale commercialization. The reason is that GRU still has problems such as too many model parameters, too long training time, and not robust enough to noisy data. Greatly limit the performance of speech recognition systems

Images