Machine Translation from Speech

This chapter describes approaches for translation from speech. Translation from speech presents two new issues. First, of course, we must recognize the speech in the source language. Although speech recognition has improved considerably over the last three decades, it is still far from being a solved problem. In the best of conditions, when the speech comes from high quality, carefully enunciated speech, on common topics (such as speech read by a trained news broadcaster), the word error rate is typically on the order of 5%. Humans can typically transcribe speech like this with less than 1% disagreement between annotators, so even this best number is still far worse than human performance. However, the task gets much harder when anything changes from this ideal condition. Some of the conditions that cause higher error rate are, if the topic is somewhat unusual, or the speakers are not reading so that their speech is more spontaneous, or if the speakers have an accent or are speaking a dialect, or if there is any acoustic degradation, such as noise or reverberation. In these cases, the word error can increase significantly to 20%, 30%, or higher. Accordingly, most of this chapter discusses techniques for improving speech recognition accuracy, while one section discusses techniques for integrating speech recognition with translation.