A Deep Learning based Approach for Identifying Spoken languages in Turkey

Spoken language identification (SLID) systems are used in areas such as human–machine interaction and security. In forensic and intelligence work, identifying the language in short or degraded recordings is important for criminal, immigration, and national security purposes. To address the specific challenges of these domains, this study focuses on extracting robust acoustic features that remain discriminative even in noisy environments and with limited temporal data. In this study, a speech dataset was constructed using languages that are widely spoken in Türkiye, including those used by foreign residents and visitors. The recordings span six languages: Turkish, Kurdish, Arabic, Persian, Russian, and German. The task of recognizing the spoken language was formulated as a multi-class classification problem. To address this task, several state-of-the-art deep learning architectures were employed, including Convolutional Neural Networks (CNNs), Transformer-based encoders, Bidirectional Long Short-Term Memory (BiLSTM) networks, and Conformer encoder models, which integrate convolutional operations with self-attention mechanisms to effectively process short utterances and complex signal patterns. These models were trained and evaluated under different data configurations to assess their robustness and generalization. Model performance was evaluated using widely accepted metrics such as accuracy, recall, and macro-averaged F1-score, with a particular focus on handling unbalanced data and phonetically similar language pairs. Experimental results reveal that the Conformer encoder demonstrated superior robustness, achieving the highest accuracy of 99.89% and a macro-averaged F1-score of 0.9914 on the population-weighted unbalanced dataset. In contrast, the Transformer model showed significant sensitivity to class imbalance, where its macro F1-score dropped to 0.8179.