An end-to-end speaker confirmation method, device and storage medium

Abstract

The present invention provides an end-to-end speaker confirmation method, device and storage medium. The method includes constructing a speaker confirmation end-to-end network, and the speaker confirmation end-to-end network includes a front-end ResCNN residual convolutional neural network model And at the back end, the ResCNN residual convolutional neural network model is used to extract speech frame-level features, and the threshold reweighted attention model converts speech frame-level features into sentence-level features, thereby completing the end-to-end network for speaker confirmation Training; the speakers obtained through training confirm the end-to-end network to determine the registrant of the test voice; the present invention realizes end-to-end processing, and the threshold reweighted attention model extracts key speech frame level features by assigning weights, Screen out non-key speech frame-level features, and then perform weighted average processing to amplify key speech frame-level features, transforming frame-level features into sentence-level features, which greatly improves speech recognition.

Description

technical field [0001] The present invention mainly relates to the technical processing field of voiceprint recognition, in particular to an end-to-end speaker confirmation method, device and storage medium. Background technique [0002] Voiceprint recognition, also known as speaker recognition, is a biometric technology, which is the process of extracting, analyzing and extracting the speaker's personality characteristics from a piece of speech, and automatically determining the speaker. Speaker identification includes speaker confirmation technology and speaker identification technology. "Speaker confirmation" refers to judging whether a passage is spoken by someone, which is a "one-to-one" problem. "Speaker identification" refers to Selecting an audio that is most similar to an unknown audio sample from the known samples is a "multiple choice" problem; however, in the current "speaker confirmation" technology, the extracted speech frame-level features are usually averaged...

AI Technical Summary

Problems solved by technology

Speaker identification includes speaker confirmation technology and speaker identification technology. "Speaker confirmation" refers to judging whether a passage is spoken by someone, which is a "one-to-one" problem. "Speaker identification" refers to Selecting an audio that is most similar to an unknown audio sample from the known samples is a "multiple choice" problem; however, in the current "speaker confirmation" technology, the extracted speech frame-level features are usually averaged for processing, and the Some non-important frames in speech features are processed together, resulting in inaccurate processing results and poor recognition results

Images