Sign language is one of the most important communication tools for hearing impaired people. In this study, the recognition of two-handed dynamic words in Turkish Sign Language (TSL) was studied using the LMC device. As a result of the repetition of 26 dynamic words, which were determined by considering the similarities and differences between them, by 6 participants, two data sets were obtained by extracting two types of feature sets. By applying a three-stage strategy to these feature sets, word recognition performances are presented by considering many aspects. These stages are data regularization, feature selection and dimension reduction. By performing these stages, new datasets with less dimension were obtained. The recognition performance was tested with six different ELM networks and the results were compared. Five-fold cross-validation was used to test the validity of the proposed system and accuracy of the obtained results. According to the results obtained with a comprehensive analysis, it has been seen that the Meta-ELM classifier maintains its performance rate and gives the highest performance. At the same time it has been observed that the Meta-ELM classifier has a stable structure that offers less user intervention.