A lazy learning-based language identification from speech using MFCC-2 features

Developing an automatic speech recognition system for multilingual countries like India is a challenging task due to the fact that the people are inured to using multiple languages while talking. This makes language identification from speech an important and essential task prior to recognition of the same. In this paper a system is proposed towards language identification from multilingual speech signals. A new second level Mel frequency cepstral coefficient-based feature named MFCC-2 that handles the large and uneven dimensionality of MFCC has been used to characterize languages in the thick of English, Bangla and Hindi. The system has been tested with recordings of as many as 12,000 utterances of numerals and 41,884 clips extracted from YouTube videos considering background music, data from multiple environments, avoidance of noise suppression and use of keywords from different languages in a single phrase. The highest and average accuracies (for Top-3 classifiers from a pool of nine classifiers) of 98.09% and 95.54%, respectively were achieved for YouTube data.