<p>Recent speaker verification (SV) systems rely on the combination of convolutional and recurrent networks to capture local and global information from input speech signals. While recurrent-based models such as LSTMs (Long Short Term Memory) can effectively extract long sequential information, they introduce high parameter complexity to the model. In this paper, we introduce a SV architecture called Bi-Directional Multi-scale K-Neighbor Residual Convolutional (BiMKRC), which combines multi-scale processing with a novel bidirectional K-neighbor learning. This approach enables each sub-feature to aggregate information from its neighboring sub-features in a forward and backward manner. This mechanism replaces the need of sequential modeling to extract global features, and reduces model complexity in terms of parameters number. We integrated attentive statistical pooling and a feed-forward network in BiMKRC architecture to generate speaker embeddings from the Mel-filterbank input features. The proposed method is evaluated on the VoxCeleb1-O dataset and compared to state-of-the-art baseline methods via Equal Error Rate (EER), minimum Detection Cost Function (minDCF) and model parameters number. Results show that BiMKRC achieves competitive performance in terms of EER and minDCF criteria, while using substantially fewer parameters.</p>