Research on Language Recognition Algorithms Based on Segment-Level Features and Automatic Identification

Language recognition technology plays a crucial role in automated speech processing within multilingual environments, particularly under the globalized context where the increasing linguistic diversity poses higher demands on language recognition systems. Due to their robustness and efficiency, segment-level features and automatic identification algorithms have gradually become focal points of research. This paper explores a variety of language signal feature parameters, including RASTA-PLP parameters, MFCC parameters, RASTA-MFCC parameters, and CMS-MFCC parameters, as well as extraction methods for sliding difference cepstrum parameters and their applications in language recognition. Gaussian Mixture Models, along with their improved Expectation-Maximization algorithms and greedy Expectation-Maximization algorithms, are thoroughly analyzed to enhance the accuracy and robustness of language recognition. Preliminary experiments verify the performance of different feature parameters and algorithms on actual datasets, demonstrating that RASTA and CMS techniques significantly enhance the MFCC parameters.