Learning segmentations that balance latency versus quality in spoken language translation.

Segmentation of the incoming speech stream and translating segments incrementally is a commonly used technique that improves latency in spoken language translation. Previous work (Oda et al. 2014) [1] has explored creating training data for segmentation by finding segments that maximize translation quality with a user-defined bound on segment length. In this work, we provide a new algorithm, using Pareto-optimality, for finding good segment boundaries that can balance the trade-off between latency versus translation quality. We compare against the state-of-the-art greedy algorithm from (Oda et al. 2014) [1]. Our experimental results show that we can improve latency by up to 12% without harming the BLEU score for the same average segment length. Another benefit is that for any segment size, Paretooptimal segments maximize latency and translation quality.