Speech Recognition, Machine Translation, and Speech Translation—A Unified Discriminative Learning Paradigm [Lecture Notes]

In the past two decades, significant progress has been made in automatic speech recognition (ASR) [2], [9] and statistical machine translation (MT) [12]. Despite some conspicuous differences, many problems in ASR and MT are closely related and techniques in the two fields can be successfully cross-pollinated. In this lecture note, we elaborate on the fundamental connections between ASR and MT, and show that the unified ASR discriminative training paradigm recently developed and presented in [7] can be extended to train MT models in the same spirit.