Voiceprint open-set identification method with unknown category internal division capability

Abstract

The invention discloses a voiceprint open-set identification method with unknown category internal division capability. The voiceprint open-set identification method has relatively high accuracy and relatively good applicability. The method is realized through the following calculation scheme: taking voice data of different speakers as input of a voiceprint open set recognition system, calculating Mel-frequency cepstrum coefficient characteristics of the voice data one by one, and training an audio coding module GE2E based on a time sequence by adopting a generalized end-to-end loss function, thereby effectively eliminating ambiguity between the speakers; through audio data coding output by GE2E, in combination with a multivariate Gaussian model and a training probability step model CGDL, judging whether any audio data belongs to a known category or not, and classifying the audio data judged to be the known category; for the audio data which is judged by the CGDL as an unknown category, constructing and training an unbounded staggered state neural network, and clustering the audio data online, wherein the obtained clustering number is the number of speakers, and all the audio data in a certain cluster belong to the same speaker.

Description

technical field [0001] The invention belongs to the technical field of voiceprint open-set recognition, and in particular relates to a voiceprint open-set recognition method capable of internally dividing unknown categories. Background technique [0002] With the development of information technology, people have more and more demands for identification technology, and identification plays an increasingly important role in the field of information security, and the requirements for its security and reliability are becoming more and more stringent. The identification technology based on traditional password authentication has exposed many shortcomings in the actual information network application, while the identification technology based on biometric identification has shown great advantages due to its unique stability, uniqueness and convenience. It has become an important research direction in the field of identification. Voiceprint Recognition (VPR), also known as Speake...

AI Technical Summary

Problems solved by technology

Although these methods have a good ability to identify unknown classes, they cannot classify known classes.

Furthermore, for speech data of unknown class, it is required to determine the number of speakers and their speaker affiliation of the speech data, which is an unsupervised online clustering problem

Because the number of speakers cannot be determined in advance, and the clustering process requires real-time performance, methods such as K-Means and LDA cannot be used

Traditional methods can use hierarchical clustering (Hierarchical Clustering), spectral clustering (Spectral Clustering) and GMM, etc. These clustering methods are unsupervised clustering processes, and it is difficult to improve the clustering effect through training

Images