Spoken language evaluation method based on deep learning and spoken language evaluation system

Abstract

The invention discloses a spoken language evaluation method based on deep learning and a spoken language evaluation system. The method provided by the invention is characterized in that voice segment intonation degree can be evaluated by adopting the deep learning algorithm, and the evaluation of the intonation accuracy of the tested voice can be acquired; the voice emotion degree can be evaluated by adopting the deep learning algorithm, and the evaluation of the emotion accuracy of the tested voice can be acquired; and the overall evaluation of the pronunciation quality of the whole sentence can be carried out by adopting the deep learning algorithm. By establishing the deep belief network mode, the DBN(deep belief network) model can be used for the spoken English test, and the evaluation of the pronunciation of the spoken English can be more comprehensive and more accurate, and at the same time, the deep learning algorithm has the higher evaluation accuracy by comparing with the emotion evaluation of the shallow model.

Description

technical field [0001] The invention relates to the technical field of speech recognition and evaluation, in particular to a method and system for oral evaluation based on deep learning. Background technique [0002] Speech signal processing technology is an important branch in the field of speech processing and speech recognition, and it is also the main core technology of today's speech recognition and speech evaluation systems. Nowadays, with the rapid development of science and technology, speech signal processing technology has penetrated into various fields, including language learning and oral automatic scoring. In language learning and automatic scoring, the purpose of using speech signal processing is to integrate the latest speech technology with current teaching and learning. Combining methods to establish a system for assisting language learning or an intelligent scoring system for spoken language. [0003] In recent years, especially since 2009, with the help o...

AI Technical Summary

Problems solved by technology

[0013] (1) DTW, but because there is no effective framework for training with statistical methods, it is not easy to apply various knowledge of the bottom and top layers to the speech recognition algorithm. Big flaw in identifying issues

[0014] (2) HMM also has certain limitations

Even so, it is still difficult to achieve good results in the recognition of non-native speech by the adaptive HMM

[0015] (3) The theoretical analysis of ANN is more difficult and cannot explain the time dynamic characteristics of speech signals well; it is easier to overfit when training and learning network models, and it is more difficult to adjust model parameters, which requires a lot of experience and skills, and The speed is slow, and the effect is not better than other methods when there are fewer layers (less than or equal to 3). Therefore, the shallow artificial neural network has not made too much breakthrough and development during this period.

Images