Automatic voiceprint modeling warehousing method, device and equipment

Abstract

The invention discloses an automatic voiceprint modeling warehousing method, device and equipment. The concept of the invention lies in that a set of data cutting thought for massive corpora is provided for the demand of speaker voiceprint modeling, so that thematic and hierarchical properties of the massive corpora are realized, and the task of voiceprint modeling and storage by using the massivecorpora is completed. The method specifically comprises the steps of performing initial filtering and distinguishing on mass data based on multi-dimensional information by utilizing a reduction thought, then performing staged progressive classification and purification operation on simplified corpus data by adopting a multi-stage superimposed consensus clustering thought, and finally obtaining corpora capable of being used for voiceprint modeling. According to the invention, a large amount of labor cost does not need to be invested for labeling, and the problem that voiceprint modeling storage cannot be achieved finally due to the fact that the error rate is possibly downloaded step by step due to direct indiscriminate voiceprint collision clustering on mass data, and the purity of modeling corpora is affected is avoided.

Description

technical field [0001] The invention relates to the technical field of speech processing, in particular to a method, device and equipment for automatic voiceprint modeling and storage. Background technique [0002] Voiceprint modeling through known target corpus data is a very important link in speaker recognition technology. Generally speaking, it is necessary to model the voiceprint of the target speaker based on the corpus of the target speaker and the voiceprint recognition algorithm, and store the voiceprint information of the target speaker into the voiceprint database. This process is the voiceprint Model storage. Among them, the quantity and quality of the corpus used for voiceprint modeling have a great influence on the subsequent recognition performance, and it is particularly important to provide a sufficient quantity and quality of modeling corpus. [0003] Therefore, it is necessary to prepare the speaker's clean corpus in advance in the process of voiceprint ...

AI Technical Summary

Problems solved by technology

However, it is more difficult to mark the speaker in the corpus (that is, to classify the corpus) compared to the recognition directions such as language recognition and continuous speech recognition.

At present, even in the case of knowing the corpus information such as the speaker and background-related knowledge, there will still be a certain error rate when labeling a small amount of corpus, not to mention when faced with unfamiliar speakers, unknown background, and unrestricted scenes. When large corpus massive data sets (the present invention abbreviates this large corpus without subject), it is difficult to classify different speakers and realize the corresponding voiceprint modeling storage operation, which is also the promotion, One of the key obstacles to the application of speaker recognition technology

Images