Analyzing the Impact of Synthetic Speech on Spoken Language Identification

Abstract

This research explores how synthetic speech influences the performance of spoken language identification systems by examining various feature types (acoustic, temporal, and rhythmic) across multiple machine learning and deep learning architectures. The study utilized Multi-Layer Perceptron (MLP), Support Vector Machines (SVM), and Long Short-Term Memory networks (LSTM) to evaluate three distinct scenarios: identifying languages from natural speech, synthetic speech, and a mix of both. Furthermore, the study investigated whether integrating all feature types could enhance system performance. The results revealed that the Mel spectrogram consistently emerged as the most effective feature across all tested models, with MLP and LSTM achieving the best overall results. In fact, the Mel spectrogram attained a remarkable accuracy rate of 100%, establishing itself as the top-performing feature. Similarly, MFCC also reached 100% accuracy in the synthetic speech scenario, highlighting its strength as the second most effective feature. Notably, combining all features did not always lead to performance improvements, underscoring the importance of strategic feature selection. The study also tackled challenges like variability in natural speech recordings and imbalances in dataset distribution, emphasizing the necessity of robust data augmentation methods. By shedding light on the interactions between feature types, model architectures, and speech data sources, this research advances the development of more accurate and resilient spoken language identification systems.