Single-versus-short speaker recognition using depth neural networks

摘要

The present invention relates to a deep network (DNN) with a triple network architecture suitable for performing speaker recognition. In particular, DNN includes three feed-forward neural networks trained according to the batching process using a cohort set of negative training samples. After each batch of training samples is processed, the DNN can be trained according to the loss function, using the cosine similarity between each sample, for example, with positive margins and negative margins to provide a robust representation of the gates. .