Enhancing Speech Recognition in the Pepper Humanoid Robot Using Deep Learning

This paper focuses on speaker recognition methods, especially GMM-based methods combined with i-vectors. In addition, the theory of methods using deep neural networks (DNNs) is also discussed. A practical implementation is made for the GMM-i-vector system integrated with PLDA, using the KALDI tool, a powerful toolkit for speaker and speech processing tasks. The experimental section examines how the number and arrangement of speakers affect the results and voices in the testing sets, and training which are performed on the GLOBALPHONE database of various languages. Results reveal that as the number of reference and testing voices increases, the recognition and verification accuracy significantly improves. This system also lays the foundation for developing DNN-based systems, in which neural networks are often used for feature extraction, and then i-vectors are applied for recognition. The final product is a standard KALDI structured script, ready to be extended and applied in future DNN systems.