Meanwhile, speaker clustering method for deep representation learning and speaker category estimation is optimized

Abstract

The invention discloses a speaker clustering method for simultaneously optimizing deep representation learning and speaker category estimation, and the method comprises the following steps: carrying out the preprocessing of a clustering voice sample, extracting I-vector features, training a convolution self-coding network, and extracting deep representation features; constructing an initial classaccording to the depth representation features to obtain a class number and an initial class label; adding a full connection layer and a Softmax layer to an encoder output layer of the convolutional self-encoding network to form a joint optimization framework, and using the Softmax layer for estimating the category of a speaker; and taking the sum of the reconstruction error of the convolutional self-encoding network and the speaker category estimation cross entropy error of the Softmax layer as a target function, and iteratively updating the joint optimization framework parameters until a convergence condition is met to obtain a voice sample of each speaker. According to the method, the optimized depth representation features and the speaker clustering result can be obtained at the same time, and the speaker clustering effect better than that of a traditional method is obtained.

Description

technical field [0001] The invention relates to the technical field of speaker clustering and voiceprint recognition, in particular to a speaker clustering method that simultaneously optimizes deep representation learning and speaker category estimation. Background technique [0002] In recent years, with the development of deep learning technology, voiceprint recognition technology has made great progress. From traditional I-vector features to d-vector and x-vector features based on deep feature transformation, voiceprint recognition has entered practical applications from theoretical research, such as network bank identity authentication, criminal detection, robot voiceprint wake-up, equipment Voiceprint unlocking and more. [0003] But training a large-scale voiceprint recognition model not only requires sufficient training data, but also needs to know which speaker each sample corresponds to. In practical applications, the training data may come from telephone recordin...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to solve the following shortcomings in the existing speaker clustering method: the feature extraction step and the speaker clustering step are carried out independently, the extracted features are not friendly to the clustering algorithm, and better clustering cannot be obtained As a result, taking advantage of the superiority of deep convolutional autoencoder networks to extract features, a speaker clustering method that optimizes both deep representation learning and speaker category estimation is provided

Images