eduzhai > Applied Sciences > Engineering >

UWSpeech Speech to Speech Translation for Unwritten Languages

  • Save

... pages left unread,continue reading

Document pages: 9 pages

Abstract: Existing speech to speech translation systems heavily rely on the text oftarget language: they usually translate source language either to target textand then synthesize target speech from text, or directly to target speech withtarget text for auxiliary training. However, those methods cannot be applied tounwritten target languages, which have no written text or phoneme available. Inthis paper, we develop a translation system for unwritten languages, named asUWSpeech, which converts target unwritten speech into discrete tokens with aconverter, and then translates source-language speech into target discretetokens with a translator, and finally synthesizes target speech from targetdiscrete tokens with an inverter. We propose a method called XL-VAE, whichenhances vector quantized variational autoencoder (VQ-VAE) with cross-lingual(XL) speech recognition, to train the converter and inverter of UWSpeechjointly. Experiments on Fisher Spanish-English conversation translation datasetshow that UWSpeech outperforms direct translation and VQ-VAE baseline by about16 and 10 BLEU points respectively, which demonstrate the advantages andpotentials of UWSpeech.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×