Meta-learning for speaker recognition in practical scenarios

Abstract

The goal of this thesis is text-independent speaker recognition where utterances come from 'in the wild' scenarios and may contain insufficient or irrelevant information. In other words, we can consider the speaker recognition for practical scenarios as long-short utterance pair and pair-matching problem. To this end, we first introduce a meta-learning framework for imbalance length pairs. Specifically, we use a Prototypical Networks and train the model with a support set of long utterances and a query set of short utterances of varying lengths. Further, since optimizing only for the classes in the given episode may be insufficient for learning discriminative embeddings for unseen classes, we additionally enforce the model to classify both the support and the query set against the entire set of classes in the training set. By combining these two learning schemes, our model outperforms existing state-of-the-art speaker verification models learned with a standard supervised learning framework on short utterance(1-2 seconds) on the VoxCeleb datasets. We also validate our proposed model for unseen speaker identification, on which it also achieves significant performance gains over the existing approaches. Secondly, for the pair-matching problem of speaker verification, we propose Cross Attentive Pooling(CAP) that utilizes the context information across the reference-query pair to generate utterance-level embedding that contains the most discriminative information for the pair matching problem. Experiments are performed on the VoxCeleb datasets in which our method outperforms comparable pooling strategies.