Neural network language models (NNLMs) have significantly enhanced speaker recognition, especially LSTM-RNN (long short-term memory-recurrent neural network). A dense architecture makes these models computationally heavyweights, so they cannot be deployed in real time or on edge with background noise. As a solution, we propose a novel method incorporating sparse representation-based LSTM-RNNs that are explicitly tuned to adapt to multi-speaker verification. We enhance hidden layer representations by incorporating sparse coding techniques, allowing for more selective and efficient neural activations. This approach emphasizes critical speaker-specific features while reducing redundancy, enabling the network to more accurately differentiate between speakers. Our sparse LSTM-RNN model retains the temporal dependencies essential for speaker recognition and significantly lowers computational complexity. Simulated results show that the model is 97.86% accurate across diverse samples of speakers, despite noisy environments. These results highlight the potential of sparse representations to optimize neural networks for speaker recognition, offering a scalable solution for real-time applications.