New Method for Automatic Recognition of Mexican Indigenous Languages: Comparative Performance of Classifiers

This work proposes a method for the automatic recognition of the Mexican indigenous languages (MILs): Mayan, Mixtec, Zapotec, Mixe, Nahuatl, Tarahumara, Mazahua, Tseltal, Chichimeco and Huichol. The long-term average spectrum (LTAS) is used as a feature for the language recognition process. In addition, the performance of classifiers such as multi-layer perceptron, sequential minimal optimization, naive Bayes, Simple Logistic and Logistic Model Tree is also highlighted. To reduce the features of the speech vector, the LTAS sequences extracted from the audios are first passed to BestFirst filters. In our experiments, high performance for MILs recognition was achieved using a simplified speech coding scheme with feature vectors with a low number of values. Our method is notable for its simplicity and efficiency, since it eliminates untested languages from the speech process. Different classifiers and tunning of its parameters were experimented with increase on accuracy of MILs recognition.