Pronunciation improvement software : An implementation of a user interface that interacts with analgorithm that uses the output from a neural network to evaluatespeech

Sammanfattning: Recently, the usage of computer software as a tool to improve pronunciation has been provento give good results compared to conventional methods. Furthermore, teachers who have beentrained to use pronunciation improvement software have had a positive attitude towards it and theyhave considered it to have high potential to improve the teaching environment. In this thesis, it isdescribed how a deep convolutional recurrent neural network (CRNN) can be created to predictphonemes from an audio file. The neural network was trained on the LibriSpeech dataset anda phoneme error rate (PER) of 3.66% was achieved for the validation set. An algorithm that isdenoted Quicksearch is introduced in this thesis that takes the output from the neural network asinput to evaluate which phonemes are pronounced satisfactorily. The Django framework is usedto create a server that is rendering a website which the user can interact with. The user will uploada recording to the Django server for analysis which is performed by the Quicksearch algorithm. It has been concluded that the neural network and Quicksearch has performance that is good enoughfor the purpose. Furthermore, it has been concluded that the Django framework has enough flexibilityto create a server with the necessary functionality. Overall, it seems as if the approach in thisthesis can be used to create software that gives relevant feedback to the user.