Context-aware Language Modeling for Conversational Speech Translation

Context plays a critical role in the understanding of language, especially conversational speech. However, few approaches exist to utilize the external contextual knowledge which is readily available to practical speech translation systems deployed in the field. In this work, we propose a novel framework to integrate context in the language models used for conversational speech translation. The proposed approach takes into account the contextual distance between a test utterance and the training corpus, a measure obtained from the external context in which the utterances were spoken. Language model probabilities are adjusted through a sentence-level weighting scheme based on this context-distance measure. When incorporated into our English-Iraqi Arabic speech-to-speech translation system, the proposed approach obtains improvements in both speech recognition accuracy and translation quality compared to the baseline system.