ThispaperintroducesWaveNet,adeepneuralnetworkforgeneratingrawaudiowaveforms.Themodelisfullyprobabilisticandautoregressive,withthepredic-tivedistributionforeachaudiosampleconditionedonallpreviousones;nonethe-lessweshowthatitcanbeefficientlytrainedondatawithtensofthousandsofsamplespe