External corpus speech identification method based on deep convolutional neural network

Abstract

The invention discloses an external corpus speech identification method based on the deep convolutional neural network. The method comprises steps that S1, speech annotation data and an Internet corpus are obtained; S2, speech signal data is cleaned through utilizing the average energy of the speech signal data; S3, feature extraction and normalization of the speech annotation data are performed;S4, a neural network model is constructed; and S5, to-be-tested speech data is inputted into the constructed neural network model, and identification text data is outputted after identification is completed. The method is advantaged in that the deep convolutional conditional random field model can be constructed based on the speech signal data, compared with a general deep learning model, fewer markup speech data are required, moreover, the cheap large-scale non-markup Internet corpus database is fully utilized to enhance the overall identification rate of sentences, sentence identification accuracy is improved, two processes are integrated, and the end-to-end speech identification method is achieved.

Description

technical field [0001] The invention relates to the field of speech signal processing, in particular to an external corpus speech recognition method based on a deep convolutional neural network. Background technique [0002] At present, there are two main categories of speech recognition methods: end-to-end speech recognition, and two-stage recognition models based on acoustic models and language models. Among them, end-to-end speech recognition is trained based on large-scale speech annotation data, the input is a speech signal, and the output is the text corresponding to the speech. The advantage of this method is that it is an end-to-end recognition system that does not require human experts to design business rules, making full use of the advantages of large-scale data and the feature learning capabilities of deep models. However, the disadvantage is that it requires huge training data support. Such methods often require tens of thousands of hours of speech data, and di...

AI Technical Summary

Problems solved by technology

However, the disadvantage is that the two-stage recognition model based on the acoustic model and the language model is a series of two-stage separate training, so that the language model does not use the original speech signal features, and cannot obtain an optimal solution.

Images