Ensemble Learners for Identification of Spoken Languages using Mel Frequency Cepstral Coefficients
Dilip Singh Sisodia, Saragadam Nikhil, G. Sai Kiran, P. Sathvik
Abstract
The most common issue in translating spoken sentences into text is first to accurately identify the speaker's language. Automated speaker language identification is useful in many applications, such as international call centers. In this paper, ensemble learning models are assessed to classify spoken languages. The Mel-frequency cepstral coefficients (MFCC) are used as primary features to retrieve necessary information from the speaker's audio samples. The delta features of MFCC's are also used for taking the time-variant behavior of cepstral. The first- and second-time derivatives of the cepstral coefficients are known as delta features. Ensemble learner models are designed using bagging, Adaboosting, random forests, gradient boosting and extra trees. All experiments conducted using ten-fold validation. The spoken language classification model's performance is compared using F1-measures, precision, recall and accuracy metrics. From the evaluation of performance metrics, extra trees classifier gives better results.