Domain-invariant feature learning method for short utterance speaker recognition

In text-independent speaker recognition tasks, performance significantly degrades as speech duration decreases, particularly when the speech is shorter than 2 s. To address this challenge, this paper proposes a domain-invariant feature learning method for short utterance. In the feature input stage, the mel-spectrogram is fused with the chromagram to provide richer input information for short utterance. In the training stage, a domain adversarial neural network is introduced to learn domain-invariant speaker features by performing adversarial training between the speaker feature extractor and the domain discriminator related to speech duration. Experiments on the Chinese speaker recognition datasets King-ASR-459 and CN-Celeb show that the proposed method effectively improves the speaker recognition performance on short utterance.