Estimating performance of pipelined spoken language translation systems

Most spoken language translation systems developed to date rely on a pipelined architecture, in which the main stages are speech recognition, linguistic analysis, transfer, generation and speech synthesis. When making projections of error rates for systems of this kind, it is natural to assume that the error rates for the individual components are independent, making the system accuracy the product of the component accuracies. The paper reports experiments carried out using the SRI-SICSTelia Research Spoken Language Translator and a 1000-utterance sample of unseen data. The results suggest that the naive performance model leads to serious overestimates of system error rates, since there are in fact strong dependencies between the components. Predicting the system error rate on the independence assumption by simple multiplication resulted in a 16% proportional overestimate for all utterances, and a 19% overestimate when only utterances of length 1-10 words were considered.