Teager Energy Cepstral Coefficients for Spoken Language Identification

Spoken Language Identification (SLID) is a key component in audio processing that facilitates the recognition and understanding of audio clips of spoken languages. Various applications, such as automatic speech recognition, multilingual voice assistants, and real-time translation services need SLID capabilities. Effective SLID enhances the transcription and processing of spoken language data and improves user experiences by providing personalized and relevant responses in the correct language. This study propose Teager Energy Cepstral Coefficients (TECC) features to capture the characteristics of spoken language. To evaluate how well TECC performs for SLID, we conducted a comparative analysis of its performance across various spectral features, namely, Mel Frequency Cepstral Coefficients (MFCC), and Linear Frequency Cepstral Coefficients (LFCC). Two classifiers, deep residual networks (ResNet-50) and Time Delay Neural Networks (TDNN), were employed in this comparison. To maximize the performance of SLID system, we applied feature-level fusion and score-level Fusion techniques to advance the state-of-the-art. Additionally, latency analysis assesses time optimization to ensure the system operates efficiently. We obtained 98.25 % accuracy in this study for two languages, and 84.25 % on 10 different languages using TECC features.