home..

Machine Translation for Conversational Speech

Modern neural approaches have greatly improved both machine translation and speech recognition. It is also now possible to build end-to-end models that directly translate speech without producing an intermediate transcription in the source language, however, this typically requires large amounts of paired data. This paired data may exist for read and broadcast speech, but it is rare for conversational speech.

When building MT systems for conversational speech, in-domain data is difficult to obtain. We are focused on improving the current state-of-the-art in the translation of conversational speech when relevant in-domain data is unavailable.

Comments? Send me an email.
© 2023 William Hartmann   •  Theme  Moonwalk