Spoken Language Identification of Indian Languages in Adversarial Synthetic and Noisy Attacking Environments

During the COVID19 epidemic, the need for contact-less biometric ASV systems is at an all-time high. As a result, voice technology based ASV systems are in more demand. In current age of Artificial Intelligence, a variety of spoofing attacks poses a significant front end threat to these systems. Such units are also constantly exposed to noise settings. In these adversarial environments, a spoken language identification module in multi-lingual systems should provide better performance. As a result, the goal of this research is to resolve the ambiguity in identifying spoken language in noisy and synthetic voice spoofing attacks. In this paper, fused CQCC-MFCC feature set is combined with a Convolutional Neural Network(CNN) for increasing performance of language detection in artificial voice attack in a noisy environment. Nine Indian languages considered here are Bengali,Gujarati,Hindi,Malayalam, Manipuri,Odia,Rajasthani,Tamil and Telugu. It was found to have a 97 percent accuracy on the INDIC TTS Database. The use of CQCC in tandem with MFCC improves accuracy by 1 % as compared to using only MFCC features.