Spoken language evaluation method and device

Abstract

The invention provides a spoken language evaluation method and device. The spoken language evaluation method comprises the steps: obtaining a to-be-evaluated audio and an evaluation text correspondingto the to-be-evaluated audio; determining an attribute characteristic value of each phoneme in the evaluation text and a posterior probability corresponding to each phoneme based on the to-be-evaluated audio and the evaluation text; extracting a pronunciation characteristic value corresponding to the evaluation text based on the evaluation text and the posterior probability corresponding to eachphoneme; generating a characteristic vector corresponding to each phoneme according to the attribute feature value and the pronunciation feature value of each phoneme; and inputting the characteristicvector corresponding to each phoneme into a spoken language evaluation model to obtain an evaluation result output by the spoken language evaluation model. According to the spoken language evaluationmethod provided by the invention, the pronunciation characteristic value corresponding to each phoneme is introduced, and the potential error of the current pronunciation can be accurately explored.Multi-dimensional characteristic information is provided for a spoken language evaluation model, and the judgment accuracy of initial consonants, final consonants and tones is improved.

Description

technical field [0001] The present application relates to the technical field of language recognition, and in particular, to a spoken language evaluation method and apparatus, a computing device, and a computer-readable storage medium. Background technique [0002] Spoken language assessment refers to the overall evaluation of a person's speaking level from different dimensions by using speech recognition technology given text and audio. Features such as the Goodness of Pronunciation of each phoneme are used to judge the speaking level of the speaker according to these features. [0003] With the development of Natural Language Processing (NLP), the Deep Neural Network (DNN) model can convert audio into frame-level phoneme pronunciation posterior probability. Based on this, the neural network-based pronunciation goodness (Neural Network-Goodness of Pronunciation, GOP-NN) is defined as the log-phoneme posterior probability ratio between the standard phoneme and the phoneme w...

AI Technical Summary

Problems solved by technology

[0003] With the development of natural language processing (Natural Language Processing, NLP), the deep neural network model (Deep Neural Network, DNN) can convert audio into the posterior probability of phoneme pronunciation at the frame level. Based on this, the good pronunciation degree based on neural network (Neural Network-Goodness of Pronunciation, GOP-NN) is defined as the logarithmic phoneme posterior probability ratio between the standard phoneme and the phoneme with the highest posterior probability. The closer the ratio is to 1, the closer the GOP is to 0. Indicates that the pronunciation is better, but in the oral language evaluation of the prior art, only the alignment information of the phoneme and the good pronunciation are used, and the information of the whole word is not used. The problem that the same standard does not apply to the innate differences of different phonemes will lead to the problems of different judgments of different phonemes by the existing oral language evaluation system, large gaps in initial and final scores, and low accuracy of tone discrimination.

Images