DNN Based Speaker Identification System Under Multi-Variability Speech Conditions
Banala Saritha, Sachin Kumar Sharma, Tungala Thiru Venkata Naga Manoj, Anish Monsley Kirupakaran, Rabul Hussain Laskar, Madhuchhanda Choudhury · 2022
Speech is the natural source of information for human identification in most biometrics, forensics, and access control systems. Mismatch in speech data is one of the biggest challenges preventing speaker identification systems from being employed in real-world scenarios. This research explores how speaker identification tasks are affected by degraded speech. Our preliminary investigation into mismatch effects in training and test conditions utilizing the IIT-G database includes mismatch in sensor and speaking styles. Convolutional neural networks (CNNs) have surpassed traditional techniques in speaker identification (SI) systems in recent years. This paper proposes a novel architecture based on a VGG-like network for an end-to-end speaker identification system. The proposed architecture outperforms the statistical methods with an improvement of 8% accuracy in identifying the speakers. The results show that the suggested approach is more accurate than state-of-the-art speaker identification techniques and notable performance deterioration compared to the matched scenario.