Application of Channel Attention for Speaker Recognition in the Wild

The objective of this paper is to build a speaker recognition system 'in the wild' (utterances with different lengths and irrelevant signals). The key elements of designing the deep neural network for this task are the type of backbone (frame-level) network, the time aggregation (utterance-level) method and the loss function (optimisation). We propose an effective speaker recognition system based on deep neural network, using SE-ResNet to extract speaker frame-level features, and a dictionary based NetVLAD or GhostVLAD to aggregate features along the time domain. We also point out that the superiority of NetVlAD plus SE-Block is that they are all based on channel attention. Additionally, we used prototypical networks loss, which learns a metric space in which the open-set classification task can be implemented by calculating the distance to the prototype representation of each class (the training process is consistent with the test scenario). We also study the influence of utterance length on the network and conclude that longer length is beneficial for "in the wild" data. Furthermore, we present results that suggest adapting from a model trained with English dataset can work on Mandarin speaker recognition, that is to say, the representations learned by our systems transfer well across different languages.