Language to language Translation using GRU method

Abstract

Current state of the art translation systems for speech to speech rely heavily on a text representation for the interpretation. By transcoding speech to text we lose important information about the characteristics of the voice like the emotion, pitch and accent. The thesis examine the likelihood of using an GRU neural network model to translate speech to speech without the requirement of a text representation that's by translating using the raw audio data directly so as to persevere the characteristics of the voice that otherwise stray within the text transcoding a part of the interpretation process. As a part of the research we create an information set of phrases suitable for speech to speech translation tasks. The thesis leads to a signal of concept system which requires scaling the underlying deep neural network so as to figure better.