Intrinsic and extrinsic component evaluation in interactive multilingual speech applications

Abstract

Due to the steady progress in technology, together with the rapid increase of powerful mobile devices, the use of voice interfaces and other speech enabled technologies has invaded our every day lives. Today people talk to their mobile phones, get directions from a car navigation system or have an electronic book read out to them. One of these emerging language applications is speech-to-speech (S2S) translation which takes on a fundamental problem of the economic and cultural exchange, the language barrier. The three main components of S2S translation, automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS), saw a huge leap forward in the last two decades due to a change to probabilistic methods which caused a boost in S2S translation technology. However, it is still evident that the output of these three components is far from perfect, deeming good evaluation methods essential.