Direct speech-to-speech translation with a sequence-to-sequence model