Identification of Suspect Using Transformer and Cosine Similarity Model for Forensic Voice Comparison

ABSTRACT Forensic voice comparison (FVC) is a crucial area within forensic science, focused on analyzing and comparing voice recordings to identify speakers. Traditional FVC methods often involve manual analysis and subjective interpretation, which can limit both efficiency and accuracy. In recent years, there has been increasing interest in leveraging advanced machine learning techniques to enhance the capabilities of FVC. This study explores the application of a semi‐automated approach using the Transformer model in FVC, with the goal of improving the accuracy and efficiency of speaker identification. The research utilizes pre‐processed forensic voice trace data in Free Lossless Audio Codec (.flac) format, along with suspect or mugshot voice samples collected by forensic experts. The Transformer model, specifically Wav2Vec 2.0, is employed for feature extraction and classification. The model processes these voice samples, using its encoder to extract relevant features and its decoder to classify the data based on cosine similarity. A threshold is applied to the cosine similarity index to determine the degree of match, with higher values indicating a stronger match. Experimental results demonstrate the effectiveness of this semi‐automated Transformer‐based approach for FVC. The similarity index achieved during testing ranged from 0.9970 to 0.9995, indicating the model's high accuracy in predicting speaker similarity. The identification process is further validated using performance metrics such as a confusion matrix, histogram, and scatter plot. This study highlights the potential of Transformer models to significantly advance the field of forensic voice comparison, offering more precise and reliable speaker identification.