Adversarial domain adaptive model training method and adversarial domain adaptive model

Abstract

The invention discloses an adversarial domain adaptive model training method. An adversarial domain adaptive model comprises a source domain embedding extractor, a speaker discriminator, a target domain embedding extractor and a domain discriminator. The method comprises the following steps of: S10, configuring parameters of a shared part layer between the source domain embedding extractor and thetarget domain embedding extractor; S20, inputting labeled source domain training data into the source domain embedding extractor, and inputting the output of the source domain embedding extractor into the speaker discriminator to obtain speaker loss; and S30, inputting unlabeled target domain training data into the target domain embedding extractor, and inputting the output of the target domain extractor and the output of the source domain embedding extractor into the domain discriminator to obtain Waserstein loss. In the method provided by the invention, the feature extractors of the sourcedomain and the target domain are not completely the same. Therefore, different parameter parts of the feature extractors can solve the conflict problem between a main task and a domain adversarial training task.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a training method for an adaptive domain against an adversarial domain and an adaptive model for an adversarial domain. Background technique [0002] Since the introduction of deep neural network (DNN) based speaker embeddings, the task of speaker verification, which aims to verify a user's claimed identity in their speech segments, has achieved significant improvements. Researchers have investigated different DNN architectures and different loss functions to enhance the discriminative power of DNN-based speaker embeddings. [0003] Despite the success of DNN embeddings for speaker verification, DNN training usually requires a large amount of annotated data with speaker labels. On the other hand, we know that the performance of a model trained from one domain drops dramatically when applied to a different domain with a different data distribution. Training domai...

AI Technical Summary

Problems solved by technology

[0005] The source domain data and the target domain data are generally very different. Use the same feature extractor to extract features for the source domain and targe

Images