Speech Recognition in Human Mediated Translation Scenarios

Human-mediated translation refers to situations in which a human interpreter translates between a source and a target language using either a written or a spoken representation of the source language. In this work we improve the recognition performance on the (English) speech of the human translator and, in case of a spoken source language representation, at the same time on the (Spanish) speech of the source language speaker. To do so, machine translation techniques are used within an iterative system design to translate between the source and target language resources. The used ASR and MT systems are then recursively biased towards the gained knowledge. In the case of a written source language representation we are able to reduce the word error rate of our English baseline system by 35.8% relative. In case of a spoken source language representation we are able to reduce the word error rate by 29.9% relative for English and 20.9% relative for Spanish