This is published under the terms of CC BY-NC.

Downloaded: 818 times

Abstract

This paper proposes a novel speaker-specific articulatory feature (AF) extraction model based on knowledge distillation (KD) for speaker recognition. First, an AF extractor is trained as a teacher model for extracting the AF profiles of the input speaker dataset. Next, a KD-based speaker embedding extraction method is proposed to distill the speaker-specific information from the AF profiles in the teacher model to a student model based on multi-task learning, in which the lower layers not only capture the speaker characteristics from acoustic features, but also learn the speaker-specific features from the AF profiles for robust speaker representation. Finally, speaker embeddings are extracted from the high-level layer, and the obtained speaker embeddings are further used to train a probabilistic linear discriminant analysis (PLDA) model for speaker recognition. In the experiments, speaker embedding models were trained using the VoxCeleb2 dataset and the AF extractor was trained based on the LibriSpeech dataset, and the performance was evaluated using the VoxCeleb1 dataset. The experiments showed that the proposed KD-based models outperformed the baseline models without KD. Furthermore, feature concatenation of multimodal results can further improve the performance.