On Deep Speaker Embeddings for Speaker Verification

In recent years, there has been a tremendous application spike in the field of deep neural networks (DNN), including increasing interest in automatic speaker recognition systems development. Currently, the utilization of DNN-based speaker embeddings, such as x-vectors or d-vectors, is a common way of creating speaker-specific acoustic models. In recent years, these DNN embedings have begun to replace standard i-vectors extracted by factor analysis. We evaluated the performance and training time of the developed systems utilising these three state-of-the-art approaches. The results obtained on the VoxCeleb1 evaluation set show that x-vectors outperformed both conventional i-vectors and DNN-based d-vectors solutions, however at the cost of a higher computational load. We also show that the x-vector system with attentive pooling, AM-Softmax activation and PLDA back-end gives the lowest error rate over other architectures.