Speech recognition

Currently, the research group has investigated deep neural network-based acoustic modeling technologies within the framework of the "Telemedicine" TÁMOP-4.2.2.A project. The group's researchers were among the first who analyzed the applicability of rectifier deep neural networks in speech recognition, and they obtained a new record of recognition accuracy on the widely studied TIMIT database using a special neural network that has a convolutional architecture. The various types of deep neural network algorithms were also evaluated and compared on Hungarian speech recognition tasks. The team's researchers won the "Emotion Sub-Challenge" of the "Computational Paralinguistic Challenge" of the Interspeech conference by making use of the AdaBoost algorithm.