Classification of languages spoken in Türki̇ye using deep learning methods

Today, spoken language identification systems are widely used in various domains such as human-machine interaction, robotics, multilingual communication, and security. These systems play an important role in applications like autonomous robotic interaction, multilingual customer services, and automatic translation. Moreover, identifying the language of anonymous voice recordings has become increasingly significant in fields such as security, forensic informatics, immigration profiling, and intelligence analysis. In this thesis, a dataset was created based on the languages commonly spoken in Türkiye, including those spoken by foreign nationals who have settled in or frequently visit the country. The dataset includes voice recordings in Turkish, Kurdish, Arabic, Russian, Persian, and German. The task of automatically recognizing the spoken language is addressed as a classification problem. To solve this problem, various deep learning architectures were implemented. First, Bidirectional Long Short-Term Memory networks were used due to their effectiveness in capturing the temporal patterns of spectral features. Additionally, Convolutional Neural Networks, which are successful in extracting spatial patterns, and Transformer-based encoder architectures, which have shown remarkable performance in recent language recognition tasks, were also evaluated. Furthermore, the Conformer encoder architecture, which combines CNN and self-attention mechanisms to capture both local and global contexts simultaneously, was applied and compared with other models. To evaluate model performance, accuracy, recall (sensitivity), and other relevant performance metrics were utilized.