End-to-end intelligent voice reading evaluation method

Abstract

The invention discloses an end-to-end intelligent voice reading evaluation method, which includes: collecting the data processing flow of reading voice, target pronunciation, error code, and speaker information used for training and evaluating the neural network; Auxiliary training; according to the target pronunciation to be evaluated and the pronunciation to be evaluated, the evaluation results are directly output from the neural network end-to-end, from the input of the target pronunciation and the pronunciation to be evaluated to the output of the evaluation results, the entire process can be differentiated, and can be directly targeted at the evaluation indicators optimization. The present invention directly constructs an end-to-end evaluation mode in which the input is speech and text to be evaluated, and the output is feedback results. It can be jointly trained with each module in the method to make the overall effect better. In addition, the auxiliary tasks constructed by the method can more accurately extract Evaluate related features to make feedback evaluation results more accurate.

Description

technical field [0001] The present invention relates to the technical field of voice evaluation, in particular to a voice evaluation method and system. Background technique [0002] With the development of speech recognition technology in artificial intelligence, people have begun to focus on how to use speech recognition technology in education-related scenarios. However, the traditional recognition evaluation method based on the hidden Markov chain model requires that the language to be recognized includes many stages of engineering processing, and requires more complex modeling. The traditional recognition method needs to split the word into phonemes and consider the context of the phoneme to form a triphone model, then construct the Markov chain corresponding to these triphone models, and estimate the path corresponding to the hidden Markov chain model from the sample probability, and finally get the recognition result. To improve these processes, domain experts need t...

AI Technical Summary

Problems solved by technology

However, the current technology related to speech evaluation needs to align the speech to be recognized with the text of interest, and often needs to extract features based on the hidden Markov chain model, which undoubtedly makes the deployment of the speech evaluation technology more complicated and makes the technology applications are limited

However, the current deep learning-based speech recognition (ASR) evaluation method needs to first recognize the evaluated speech, and then align the recognition result with the text to be evaluated to obtain the evaluation result, resulting in a relationship between the evaluation quality and the recognition result.

In such a case, the user's accent has a great influence on the evaluation results; on the other hand, the recognition results need to be aligned with the evaluation text, and it is impossible to achieve a complete end-to-end output error position. The recognition results can only be optimized indirectly, and cannot be directly optimized. evaluation results

Therefore, some evaluation errors that are not caused by recognition (such as pitch, rhythm, etc.) are difficult to be correctly judged based on this method of recognition

Images