Thisis a goodtechnologytohangourhatonbecauseitisgettingsomuchbetter.
これだけ良くなっているのだから、この技術には脱帽です。
Everysingleyear.
毎年、毎年。
Machinelearningisbeingbakedintoworkflowshelpingcreatepreviouslyunimaginablemomentsfrombigblockbusterstononfictiontv I thinkwhereAireallyisimpactfulisgettingittodothingsthathumanbeingscan't do, includingraisingthedeadasifyouknowyouhaveAndyWarholstandinginthestudiorightinfrontofyouandyoulookedathimandsaid I wantyoutosayitlikelikethis.
Soon a 10 dayshot, onedaywassimplyremovingtrackingmarkers, developed a neuralnetwhereweareabletoidentifythedotsonthefaceoftheartificialintelligenceaveragedouttheskintexturearoundthedotremovethedotandtheninfilledwiththeaverageofthetexturesurroundingit.
Youcouldhirelike a voiceactortodolike a greatimpersonationbutwefoundwithhisvoicesyoukindofwantedtoretainthathumannessthatAndyhadhimself, youcangetfairlyclosewiththevoiceactorbutyoureallycan't getit.
Soandthat's whereaitechnologyreallyhelpsgeneralofaudioistheabilityfor a artificialagenttobeabletoreproduce a particularvoicebutalsoreproducethestyle, thedelivery, thetoneofof a realhumanbeinganddoitinrealtime.
Andtheysentustheseaudiofiles, recordingsover a telephone.
そして、電話で録音した音声ファイルを送ってくれました。
They'reallfromthelateseventies, midseventies.
どれも70年代後半から半ばにかけてのものです。
Thethingaboutmachinelearningisthatbaddatahurts a lotmorethangooddata.
機械学習について言えば、悪いデータは良いデータよりもずっと痛いということです。
So I rememberlookingatthedatawehadavailableandthinkingthisisgoingtobereally, reallydifficulttogetrightwiththreeminutesofdata, we'rebeingaskedtoproducesixepisodesworthofcontentwiththreeminutesofhisvoice.
Although I guess I wantedtobewe'reabletosaythatlineandthenouralgorithmsareabletoextractcertainfeaturesoutofthatdeliveryandapplyittoAndy's syntheticortargetvoice.
What I seehappeningrightnowisactuallyquitetheoppositeisthatitiscreatingnewopportunitiesforustospendthetimeondoingthingsthatarecreativelymeaningfulratherthanspendinglotsoftimedoingmenialtasks, wereactuallyabletofocusonthecreativethingsandwehavemoretimeforiteration.
So, youknow, we'renotinterestedinlikecreatinganadspotthat a realvoiceactorwoulddobecauseinallhonesty, thatrealvoiceactorwoulddowaybetterthantheAitechnologywoulddowouldbewayfasterifyou'rejustdelivering a particularsentenceor a particularline.
I thinkitcomesdowntoeducatingthegeneralpopulationeventuallymakingthemunderstandthattheyshouldthinkthroughwhatevertheyare, lookingat, whateverthey'rereadingandnowwhateverthey'rehearing, wefeelwe'redirectionallycorrectinourbetthatthisis a goodtechnologytohangourhatonbecauseitisgettingsomuchbettereverysingleyearandwedon't wanttomisswhatweseeaslike a oncein a lifetimeopportunityhere.