Deep embedding learning based speaker verification (SV) methods have recently achieved significant performance improvement over traditional i-vector systems, especially for short duration utterances. Embedding learning commonly consists of three components: frame-level feature processing, utterancelevel embedding learning, and loss function to discriminate between speakers. For the learned embeddings, a back-end model (i.e., Linear Discriminant Analysis followed by Probabilistic Linear Discriminant Analysis (LDA-PLDA)) is generally applied as a similarity measure. In this paper, we propose to further improve the effectiveness of deep embedding learning methods in the following components: (1) A multi-stage aggregation strategy, exploited to hierarchically fuse time-frequency context information for effective frame-level feature processing. (2) A discriminant analysis loss is designed for end-to-end training, which aims to explicitly learn the discriminative embeddings, i.e. with small intra-speaker and large inter-speaker variances. To evaluate the effectiveness of the proposed improvements, we conduct extensive experiments on the VoxCeleb1 dataset. The results outperform state-of-the-art systems by a significant margin. It is also worth noting that the results are obtained using a simple cosine metric instead of the more complex LDA-PLDA backend scoring.

In this paper we present an effective deep embedding learning architecture, which combines a dense connection of dilated convolutional layers with a gating mechanism, for speaker verification (SV) tasks. Compared with the widely used time-delay neural network (TDNN) based architecture, two main improvements are proposed: (1) The dilated filters are designed to effectively capture time-frequency context information, then the convolutional layer outputs are utilized for effective embedding learning. Specifically, we employ the idea of the successful DenseNet to collect the context information by dense connections from each layer to every other layer in a feed-forward fashion. (2) A gating mechanism is further introduced to provide channel-wise attention by exploiting inter-dependencies across channels. Motivated by squeeze-and-excitation networks (SENet), the global time-frequency information is utilized for this feature calibration. To evaluate the proposed network architecture, we conduct extensive experiments on noisy and unconstrained SV tasks, i.e., Speaker in the Wild (SITW) and Voxceleb1. The results demonstrate state-of-the-art SV performance. Specifically, our proposed method reduces equal error rate (EER) from TDNN based method by 25% and 27% for SITW and Voxceleb1, respectively.

No abstract

Deep embedding learning based speaker verification methods have attracted significant recent research interest due to their superior performance. Existing methods mainly focus on designing frame-level feature extraction structures, utterance-level aggregation methods and loss functions to learn discriminative speaker embeddings. The scores of verification trials are then computed using cosine distance or Probabilistic Linear Discriminative Analysis (PLDA) classifiers. This paper proposes an effective speaker recognition method which is based on joint identification and verification supervisions, inspired by multi-task learning frameworks. Specifically, a deep architecture with convolutional feature extractor, attentive pooling and two classifier branches is presented. The first, an identification branch, is trained with additive margin softmax loss (AM-Softmax) to classify the speaker identities. The second, a verification branch, trains a discriminator with binary cross entropy loss (BCE) to optimize a new triplet-based mutual information. To balance the two losses during different training stages, a ramp-up/ramp-down weighting scheme is employed. Furthermore, an attentive bilinear pooling method is proposed to improve the effectiveness of embeddings. Extensive experiments have been conducted on VoxCeleb1 to evaluate the proposed method, demonstrating results that relatively reduce the equal error rate (EER) by 22% compared to the baseline system using identification supervision only.

No abstract