Identification of true speakers from disguised voices in anti-forensic scenarios using an efficient framework

Audio forensics in the criminal justice system always remains significant support. Voice prints are of great interest to investigators as partial evidence of illegal activities. However, with rapid technological advancements, people with illicit intentions often utilize different voice changers as anti-forensic weapons and disguise their voices electronically while performing unlawful activities. Several automated machine learning-based techniques are adopted in true speaker identification. Our proposed analytical framework compares traditional forensic laboratory methods with automated deep learning methods. The analysis is conducted based on the likelihood of acoustic parameters, including similarities and dissimilarities. For this, we have collected our own data set, consisting of 105 speech recordings, 35 of which were from females and 70 from males. To determine the robustness of our framework on our collected dataset, we created disguised voices using electronic and non-electronic methods. Our proposed deep learning model outperforms machine learning and automatic lab-based speaker identification systems with \(95\%\) $95\%$ test accuracy. The model also correctly identifies \(97\%\) $97\%$ of the real audio samples and \(75\%\) $75\%$ of fake ones. The results demonstrate that our CNN model is more effective in detecting disguised voices than other forensic methods in our framework. To further evaluate the performance, we also compared our proposed forensic speaker identification model with other state-of-the-art methods and found it to be superior.