Raw Waveform Based Text Independent Speaker Identification under Multivariability Speech Conditions

Speech is a vital modality for human identification in various applications such as biometrics, forensics, and access control. However, mismatched speech data remains a major challenge in deploying speaker identification systems in real-world scenarios. This research investigates the impact of degraded speech and mismatched training-testing conditions using the IITG-MV database, with a particular focus on sensor and speaking style variability. The paper proposes a convolutional neural network (CNN)-based architecture for text-independent speaker identification using raw waveforms, thereby eliminating reliance on handcrafted acoustic features. The proposed model is designed to efficiently capture speaker-specific characteristics, such as pitch, timbre, and formant structures, under multivariability conditions. Experimental results on the IIITG-MV dataset demonstrate a 10% enhancement in accuracy and a 32% decrease in validation loss relative to conventional CNN approaches. These results indicate the robustness and effectiveness of the proposed method for real-world speaker identification applications. However, the model also shows significant performance degradation in highly mismatched conditions.