An Iterative Update Method for Speech Training Data Based on Phase Test Feedback

Abstract

The invention discloses a voice training data iterative update method based on phase test feedback. The method continuously adds misrecognized voices in phase test voices for training to establish a reference model, and then calculates the likelihood score of the original training voice on the reference model. , and then sort all kinds of voices according to the likelihood score, and select the sorted voices according to a certain ratio to obtain the stage core training voice. Through the voice training data iterative update method provided by the present invention, high-quality training voice can be continuously screened according to the feedback of test data, and the obtained phase core training voice utilizes phase application feedback in time, so its future recognition performance can be improved. It will get better and better; this method is suitable for speech classification scenarios such as speech recognition, speaker recognition, and fake speech recognition.

Description

technical field [0001] The invention belongs to the technical field of speech recognition, in particular to a method for iteratively updating speech training data based on stage test feedback. Background technique [0002] As a biometric authentication method, the voiceprint authentication system has the advantages of low acquisition cost, easy access, and convenient remote authentication. It has been widely used in access control systems, financial transactions, judicial identification, and other fields. With the rapid development of speech synthesis technology, on the one hand, it has brought more convenient services and better user experience to people, such as Zhensheng intelligent customer service, Zhensheng intelligent navigation, audio books, intelligent voice calls, etc.; It brings great challenges to the security of the voiceprint authentication system, such as using synthetic voice to attack the voiceprint authentication system to make its performance significantly...

AI Technical Summary

Problems solved by technology

The existing experimental research on synthetic speech detection is trained according to the training set set by the competition, usually using a large amount of training data; however, in actual situations, when more training data is used, the performance is improved. Decline, because there is redundancy in the training data, it is necessary to perform data selection

Images