English to Arabic speech to speech system

Abstract

New English to Arabic speech-to-speech system has been developed in the thesis. The system allows rendering as speech in the Arabic target language the speech uttered in the English source language. Three components form the system. Automatic speech recognition using DeepSpeech system transcribes the uttered English speech into intermediate text. The recognized English text is translated using OpenNMT to produce the corresponding Arabic text. Tacotron2 synthesizes speech out of the Arabic text. A dedicated Arabic dataset has been collected and labeled to train the Tacotron synthesizer. The machine translation models have also been trained. The present document describes the different obstacles faced and the solutions provided. It also provides results. While the different components have state of the art performances, the first subjective tests prove the good performance of the overall system.