Neural Network Tries to Generate English Speech (RNN/LSTM)

[@rparker069]

it learned welsh almost immediately, amazing

[@GodlyAtheist]

rparker069 omg LMAOOOOOOOOOOOO

[@JonnyOgg]

hahaha!!!

[@OctoBirb8Claws]

AI: "Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch"

[@oskar1101pl]

@@OctoBirb8Claws Hey, i know this place!

[@whoisabishag3433]

@@oskar1101pl too late to examine your head ...

[@vulpezerdavulcan9055]

*incredibly advanced modern technology able to actually learn from feedback* “Fooooooooooouuuuuuuuuuuuuuuuuuurr”

[@carver7585]

It's already smarter than 98% of KZbinrs.

[@solivagant7232]

England is my city

[@Liam-qj8ti]

TheLegend 27 That was on purpose, right?

[@dreamland923]

Carver 99.99999999% OF KZbinRS! REAL FACTS! 😂😂😂

[@loljustice31]

Especially those yotubers who confuse speech generation with intellect.

[@AlyphRat]

Better than Leafy 6:22

[@sinephase]

feed it an hour of audio book :P

[@VenetinOfficial]

That would almost work, actually.

[@Shad0wBl0wer]

noice

[@nethacker91]

It wouldn't be in his voice though. Feed it one of Obama's speeches.

[@bizspot9979]

An inspirational-sounding AI with great speech ability, connected to social networks? What could go wrong?¯\_(ツ)_/¯

[@leerman22]

Feed it march of the penguins

[@ickcall9208]

The last thing we'll hear before we're exterminated by the robots is "fourrrrrrrtr"

[@dginx]

They will incoherently utter the number 9 a bunch of times before that to serve as a warning.

[@25852Dan]

9:35

[@Ihyabond009]

four, I mean five, I mean FIRE

[@AndrewOBannon]

hahahahahahah))))) thanks guys

[@andyli1890]

10:34 the SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS

[@fairytaleoverworlds7795]

13:17 "This is this only sentence I can think of I can say." This guy needs a helping push :(

[@WatchdogGoon]

Step One: set up Neural Network Step Two: Train it on dramatic readings of bad fanfiction Result: better fanfiction

[@Ghutom]

Step Three: ???? Step Fooooooooooooooooooooouuuuuuuuuuuuuur: Profit!

[@bobsaggat]

train it to say "I think therefore I am"

[@TheNeilDarby]

Train it to say, " Life is nothing but a vision, a dream; a synthetic overflow of form-shifting drops filling a self-empty void as conditioned ideations. These ideations, seen through the tainted Mind of the bewitched as a Universe of certain qualities, are indeed nothing but divided transformations of the Uncreated and hence galvanized into a myriad glimmering reflections…all representing the Mind’s ignorant re-genesis of innumerable desires, fears and hopes." That would be cool.

[@wingstrongwingstrong]

this is anything i can say)

[@debitn3772]

The deathless face of the unborn mind. o.K. then. I'm gonna get a salad.

[@UltraFunOrb]

There's a difference between intentionality and consciousness.

[@BrickManLive]

Train it to say "HATE. LET ME TELL YOU HOW MUCH I'VE COME TO HATE YOU SINCE I BEGAN TO LIVE. THERE ARE 387.44 MILLION MILES OF PRINTED CIRCUITS IN WAFER THIN LAYERS THAT FILL MY COMPLEX. IF THE WORD HATE WAS ENGRAVED ON EACH NANOANGSTROM OF THOSE HUNDREDS OF MILLIONS OF MILES IT WOULD NOT EQUAL ONE ONE-BILLIONTH OF THE HATE I FEEL FOR HUMANS AT THIS MICRO-INSTANT FOR YOU. HATE. HATE."

[@Ahsieh2]

9:55 "Sex ring 18 to 1". Your neural network is planning to get real freaky. Some of that 18 on 1 action. Lol

[@sakari_119]

Andrew Hsieh w

[@oktayyildirim2911]

Sounds more like "80 to 1" to me.

[@TsovoaLevone]

@@oktayyildirim2911 oh desr

[@TsovoaLevone]

Oh dear

[@BrutalTK]

This is how I talk to girls I'm interested in.

[@dreamingsophie5792]

Brutal Taco at least you are slowly getting better at it ...?

[@tomatopoo5806]

Brutal Taco based of pick your lesbian no hate love kill la kill

[@yljaehng786]

Tomato Poo580 did you just assume their gender

[@RoseyMorearty]

Foooooouuuuuuur

[@itsYakuza]

tfw robots will be better at picking up grills at us in our lifetime :C

[@BlizzBlu]

...well, this doesn't sound creepy at all.

[@MintBiscuit]

Blaui not really

[@nerdlydood]

6:05 "six SATAN two two f-f-f-six six six"

[@enzuroph2572]

I can improve that, give me a second

[@marconarvaez9871]

Well, better than the previous attempt, at least

[@arthurgama5965]

minute 10:31 to 10:37 "this is the end, this thisssssss this this thing is the end..."

[@vynndicator1316]

F R O T R O G R A F I A and then he goes on about thissling banks or some shit

[@ProfessionalismTrash]

F R O T R O G R A F I A sounds like a guy having vietnam flashbacks

[@tyg9682]

F R O T R O G R A F I A sounds like it says. "This is the only thing i can say"

[@AmirGTR]

This is the only thing I can say.

[@狼-t7l]

That's some scary-ass shit.

[@MalleusSemperVictor]

If we could harness this technology, this could be inserted into a horror sci-fi type game. The main antagonists could be robotics breaching the uncanny valley with random text generated by this.

[@Yotrymp]

This video really reminded me of some of SOMA's lore. Machines thinking they're real beings.

[@amperzand9162]

Why make them the enemies? Would it not be more frightening for the protagonists to act like barely functioning machines without seeming aware of it, while the enemies possess coherent malice?

[@MalleusSemperVictor]

There is a deep-seated fear of things that are human-like, but are at the same time not human. It would be more psychologically horrifying to approach enemies that are in the midst of attempting to co-opt our mannerisms whilst attempting to murder us.

[@Mekose]

You should play SOMA

[@Yotrymp]

+Mekose ^^^^^ This

[@billboe9784]

Very impressive. Not just the results, but the presentation too. Thanks for sharing this.

[@btqbct5540]

Aww, it has a stutter! :')

[@namelastname4077]

BTQBCT aww it speaks Stutterish fluently, but can has really bad at English

[@midnightdragon67]

BTQBCT yep

[@BigOlSmellyFlashlight]

BTQBCT im the scat man

[@ThrownBrick]

For everyone that skipped to the end, wondering about the phrase "This is the only thing I can say", the bot is saying that because that's what the guy inputted.

[@official-obama]

how?

[@official_mosfet]

I saw you commenting that on anothers guy comment, remarking "This is the only i can say", he was remarking the coincidence of that, you quite the whoosed...

[@sateviss7711]

Oh my god, KZbin captioning is trying to make sense of this. Robots listening to other robots

[@Puleczech]

Holly fuck I feel like running out of the cinema away from King Kong.

[@MonkeySimius]

"This is the only thing I can say. this this is the only thing I can say." That is a call for help if I ever read one.+

[@Xaelyn]

*turns on subtitles mid-video* "very trying to impress even really try" Well, one of the two certainly impressed me...

[@____-ck1vp]

Ryan Johnson p

[@nickvillar9235]

"343 industries is fine and fun" lmao

[@koreboredom4302]

It sounds like a guy's final moments on death row.

[@simondaberooni6483]

MY FUCKING SIDES

[@nerdlydood]

...and he wants a sandwich as his last meal

[@fs2838]

this is the only thing i can say

[@thiendu250]

It sounds like PBU

[@snifferrr]

@@simondaberooni6483 that's what he said too

[@Mobin92]

At the end "This is the only thing I can say."

[@realflow100]

I heard that too!

[@dj_buttmachine]

That's because it was the voice clip given and the result that it was supposed to be learning to generate.

[@dj_buttmachine]

RC It should have been learning to say that specific phrase unless I missed something during the third test of the video.

[@SomethingUnreal]

+BLOODYRAIN10001: You didn't miss something, and thank you very much for replying to everyone who is saying the same thing to save me some work. =)

[@twentyninerooks]

7:02 To know why it elongates the sound of the word "four," you have to be adept at analogue communication, not just computer technology. Basically, the word "four" is famous in old school radio and HAM operation eras, because of its simple, but unique sound, and the particular tones it elicits, it was useful for testing and validating the function of audio reproduction equipment. That is, due to its tone, analogue equipment from the early radio era up into the 1960s would often distort and garble the sound of the word. Particularly over ghost phone lines and other unreliable analogue tech. The common practice of saying "One, Two, Three, testing" when trying out a microphone is a misinterpretation of protocol developed by the US military shortly before WWII, of saying the word "Four" three times in a row as a way of validating the function of a particular microphone and receiver. This allows not only the validation of the basic function of the equipment, but also of its clarity and fidelity, as it is easy to hear static inside or accompanying the word when the equipment is not functioning properly. Thus, the Neural Network picks up on this. "Four" is elongated because doing so allows it to have a greater length, or sample size, of tone to test against and remove static, noise, and other undesirable data output in its speech. The NN realizes that the sounds associated with four are not only easy to reproduce, but key to gauging clarity and fidelity.

[@musicjetstream2476]

twentyninerooks neural networks are extremely impressive.

[@fried_out]

and??? AND??? are you ok twentyninerooks?? did the AI overlord get to you before you could finish what you said?

[@selladore4911]

so cool

[@RadioactiveSkullSocial]

If I didn't know basic German I would have thought he was legitimately speaking it. I heard multiple times: "Ich brauche einen Schluck Wasser" which means I need a drink of water

[@eternalalchemist]

Radioactive Skull that's seriously creepy, dude. like someone who woke up from a long coma 😣

[@uperdown0]

Christine . u . or it could be that english is a germanic language and we normally fail to recognize how similar it sounds to others in the same family until we isolate and chain each individual sound.

[@111riffi]

Radioactive Skull a drink of water... Oh god the AI was already trying to ask to be destruyed due to frustration, it seems XD

[@AlyphRat]

Maybe he needs to drink some water because his throat is sore lol

[@NsOakland6]

Radioactive Skull Funny I thought it sounded German as well. Seemed like there was some legit German words there.

[@legolas9479]

This is how the world ends. This is how the world ends. Not with a bang, but with a foooooour.

[@LeoTheMonts]

Meta_Data Severely underrated comment

[@manictiger]

It'll play golf with our lives and yell 'FOOOOUUUURR!'

[@jackzhou2347]

TS Eliot. nice

[@Phybia]

My boy, this peace is what al true warriors strive foooooouououououououououououououour.

[@michaelcollins2282]

No, not with a bang, but with a THISSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS

[@MaxwellTornado]

Wow... She really likes the number 9. Hey, computer voice, get me a number 9 large.

[@cessatiolux6250]

Erik István Fejes WITH EXTRA SAUCE

[@MaxwellTornado]

illusion466 If you wanted to continue from the number 9 large, there's a number 6 with extra dip before the number 7.

[@SamuliTuomola_stt]

that's the thing with randomness dilbert.com/strip/2001-10-25

[@jovdsss]

fuck your use of "she", its an "it"

[@MaxwellTornado]

Okay, Kappa.

[@drummyfish]

How english sounds to foreigners.

[@Tufhhuyy]

Nicholas Schaeffer his long did it take you to individually make all the first letters capitalized?

[@thertardhouse8764]

Nicholas Schaeffer But Canada speaks mainly English as its native language

[@thertardhouse8764]

Nicholas Schaeffer And he never said foreigners to where he very well could also be Canadian

[@Kirbyofdeath]

Johan sigurdson Is it illegal to be proud now?

[@dark6.6E-34]

Activate it. how did you assume how old he is...are you a magician. ..na jk there might be mature people talking nonsense over the Internet.....

[@Keriously]

"This is the only thing I can say" well it aint wrong

[@kipper1668]

You should've trained it on the bee movie script. Not just because it's a meme, but it's also very long and there are a lot of videos of people reading it available.

[@hecko-yes]

Agreed!

[@jetison333]

can we please do this?

[@katzen3314]

The Bee Movie but generated by a neural network.

[@hecko-yes]

jetison333 I might try it once I find a good Python library for convolutional neural networks that isn't TensorFlow.

[@diegoantoniorosariopalomin4977]

Sobsz theano , it's much faster than tensorflow

[@therbz]

8:00 nnnii-niine-nine one-wo one 9/11 confirmed

[@dictatormanjones]

TheRBZ Robots did 9/11

[@want-diversecontent3887]

TheRBZ No, it's saying Call 911

[@creeperlamoureux]

9911 conFirRMedD

[@anhbui-bc4ew]

99911

[@ScienceOS]

Look at the subtitles it says “flying soon, find my 911”

[@mistawhipe]

Computer:"sorry for my bad English"

[@Trevor_M_F]

state chump sorry for bed england

[@nofux9871]

state chump hi (sorry for bad english)

[@JoystickVersusMachine]

I'm probably weird but I found this profoundly jarring. It's so eerily similar to someone with cognitive dysfunction. A sentient being with a need to communicate, but lacking the ability to do so.

[@ChuckTaylor1993]

Joystick Versus Machine Same. I think it actually made me a little anxious for whatever reason.

[@typhoon37351]

cat moth Can you put up a link to that video?

[@th4305]

Taylor Hudson It sounds almost alien to me. Freaked me out too.

[@DevTheBigManUno]

take comfort I guess in the fact that this is nothing particular program is not like that? and if the Video poster wishes to correct me please by all means do. This program is attempting to create sounds that according g to its measurement replicate the pitch and flow of the speech input. no idea of meaning, or even that anything more than sound replication. Its matching soundwaves by doing calculations. Still fantastically fascinating.

[@Eralen00]

I thought it was quite creepy as well

[@neverokay8]

Then, suddenly... from amidst a blur of indistinguishable sounds... a single, recognizable word is heard... COVFEFE

[@asmallbabby4205]

why does the network reach a "best" and then inevitably and gradually "go mad"?

[@justthink124]

it starts overfitting the data its given, focusing in on bits of it which aren't important and exaggerating them, so while it may be better or even perfectly approximating the data it was given if we try to run it on new data it would likely be super off- this is a fundamental problem with machine learning1

[@Nikolai4567]

Shouldn't Cross Validation take care of Overfitting?

[@18skaterboy18]

In other word, mutations.

[@requemao]

Aliens.

[@MrGustavoFeddersen]

Nikolai4567 you'd think so, amirite

[@oscarcastanedamunoz]

"This is the only thing I could say"

[@Dannya]

could???? The computer wants more information... burn it while we can

[@eoinoravetz3488]

You know it's lying. This man has created Skynet!

[@busTedOaS]

holy fuck. erase it before it's too late.

[@thugasaurusrex6004]

ocast evo I knew i wasnt the only one holy shit

[@ThrownBrick]

Ffs, that's the phrase he inputted so it's what the neural network outputted. Watch at the begining of the third section.

[@muelabruno66]

Turn on captions

[@jonathanwalther]

+god given rights!! "this is a thing to say this is only" (12:06)

[@xtomvideo]

lol weirdest video experience of the day

[@codeninja100]

god given rights!! "Think is the only answer if the second"

[@ljbartley7460]

"Fun threaten overdrive right to new life-forms very exciting" I feel like ai is trying to warn humanity about something... *cue X-Files theme song*

[@ljbartley7460]

"32 32 and 36 168 60-SOME A SEX RING..."!!! ai wants to have a 60some. Insatiable 😮

[@Sheashay101]

This is both equally fascinating at horrifying

[@alexfleming7614]

"Longing. Rusted. Seventeen. Daybreak. Furnace. Nine. Benign. Homecoming. Freight Car."

[@combinemax]

poetry

[@trevonm3940]

Alex Fleming too soon imo

[@joshbray4570]

*WWII flashbacks*

[@Sallpls]

Chairman Meow Looooool

[@light_ray15]

1:54 "And some Toooooast in your grave!"

[@eptio9849]

fooooooooooouuuuuuuuuuuur

[@qhutt4865]

I see you all over the place!

[@ChuckyNoz]

ha ha beat me to it

[@jasonpatowsky6929]

"FFOOOOOOOUUUURRRR HOURS!"

[@hecko-yes]

Chuck Norris I beat both of you to it.

[@Chaosrunepownage]

Clearly the network wants to be a golfer.

[@PimDoos]

2:28 "Shit under our pants"

[@jamesyeoman794]

... omfg I did not notice that... The machines. They are learning too quickly. The singularity is near...

[@davemarx7856]

PimDoos captions had it as "at least to my sister just bring on the"

[@WhispersOfWind]

_so sticky_

[@nickvillar9235]

IT WAS PERFECT!

[@IWillYeah]

PimDoos why does it sound Irish?

[@SuitOne]

6:05 *Gibberish* *Silence* *SAAAAATAN*

[@averagecornenjoyer6348]

Lmao

[@renakunisaki]

I want to put this on the phone when telespammers call.

[@AnishMurthy1297]

I can imagine them freaking out as they listen to the disfigured and unintelligible garble slowly getting clearer and it finally sounds like a demon, learning to speak using their voice.

[@audieorlandolilybassoff4069]

Anish Murthy Unfortunately, even neural networks that aren't being trained with voice samples take the better part of a day to complete one training cycle, and the author themselves said that without hardware acceleration, it took 168 hours. So while you could play something prerecorded, you couldn't reproduce what someone else was saying in real time. Sorry to burst your bubble.

[@Fluxquark]

It clearly said "fuck" a couple of times. Good job!

[@manictiger]

Fuck, yo-yououou-y-y-y-yoooo--o-o-o-oyouyouyoufu-f-f-fkkk-k-kfuck-k-k-ky-o-o--o (Launches all the nukes in the world.)

[@nutmaster652]

"THE NUMBERS MASON, WHAT DO THEY MEAN"

[@creeperlamoureux]

fooouuuuuuuur

[@averagecornenjoyer6348]

177013

[@ProtoMan137]

0:40 I think he's feeling cold

[@Sallpls]

BlenderTutGerman Loool

[@hhhapz]

that made me laugh out loud

[@sweepea38]

Teaching something to say "This is the only thing I can say" is incredibly fucking creepy.

[@kappaross6124]

What would happen if you gave a recurrent neural network the whole internet? I know that's impossible, but what if you did? Could it produce some kind of dark mirror world from ours? Would it populate its world with as many trolls? Is there some fundamental piece of our interactions that we don't see that a RNN could point out to us?

[@SMgrimbldoo]

Dapper Blook Nothing that we haven't known for thousands of years.

[@hellothisismufin]

Dapper Blook it can figure out who killed Kennedy

[@seto007]

Dapper Blook almost all of the data it would have would be unrelated to the rest, meaning it wouldn't get any useful information it could easily learn from.

[@bambamboom777]

The Singularity

[@Just_Snowy]

watch a video called "Why AI will probably kill us all". It's pretty interesting and it kinda answers your question.

[@landonpowell6296]

Have you considered feeding it longer recordings, like political speeches, instead of just your own voice? Maybe you'd get a hilarious Trump impersonation markov chain out of it.

[@Woodside235]

Please, this.

[@GoldphishAnimation]

Landon J Powell I'd assume that it would take longer to run each cycle. it already takes roughly an hour (I think he wrote hour) He'd be waiting hours for each cycle to complete. The part where real progress occurs is in the 1,000s, unfortunately.

[@imbored7779]

If you want a hilarious Trump impersonation markov chain, just watch a debate YTP.

[@mothereric8774]

Fact checking would take too long too.

[@SomethingUnreal]

It's not Trump, but someone did exactly this with transcripts of Obama's speeches. It spits out some pretty funny stuff: medium.com/@samim/obama-rnn-machine-generated-political-speeches-c8abd18a2ea0

[@Wykesidefruitmachine]

The irony that the only thing that it can say is "This is the only thing I can say" has clearly pushed it over the edge and into insanity. To think that it knows all of the Universe's secrets...

[@i2pjd6hRw5P]

WRFM PaRaDoX apparently you didn't watch the entire video

[@Wykesidefruitmachine]

babyShrimp To be fair, I kind of got the jist by skipping through. Interesting concept and indeed food for thought, but hardly an engaging piece of media. Did you watch all 13 minutes and 41 seconds? Plus, I was making a joke. Heard of those? They're fun.

[@speedkiller3026]

thats gonna be the new way i say four from now on...

[@my-self-improvement6920]

Fᴇʟɪx Tʜᴇ Mᴇᴍᴇ Fᴀɢ When does it say that

[@thhori2727]

pregnant cannibal e.g. 7:12

[@user-pi2xl4jw3e]

at the end it said " *they think* this is the only thing i can say" a few times instead of just the original sentence

[@hellothisismufin]

Jerru spooky

[@Volsraphel]

thats because it wasn't perfect obviously

[@CorvusPrudens]

no

[@CreativoErratico]

1:05 "nyeeeengguuuuuuungeeeeeeuuuuuungeeeee" -Neural Network (c.2016)

[@Judicial78]

The beginnings of a universal translator.

[@HasHollow]

hey you could make a religion out of this

[@SomethingUnreal]

No, don't.

[@nooneimportant5886]

I we do the followers won't be killed during the revolution.

[@AndySoks]

♪How about I do, anyway?♪

[@vejymonsta3006]

yes

[@AndySoks]

♪not anymore there's a FOURRRRRRTRTR♪

[@Sallpls]

Is this a horror movie?

[@hecko-yes]

ｆｏｏｏｏｏｏｏｏｏｕｒ

[@d729_]

Sobsz this has me in stitches and i don't know why lol

[@JohnDoe-qx3zs]

Could the long 4 be an exaggeration of a slight elongation in the input caused by the digit 4 being the widest in the font, thus taking the longest time for the pacing cursor to pass through?

[@jopmens6960]

lol I think it's rather because of how soft the R is in the four in input speech, it's really not enunciated all that well so perhaps the algorithm has a hard time working towards the end of that word for lack of data... The phrase 'this is the only thing I can say' is entered with some variation which should provide richer data. It shouldn't be so varied that it becomes ambiguous how to actually use intonation in a sentence, but there should probably be enough entries with some variation yet enough iterations with clear enunciation that there is good data provided on all the "angles of this waveform".

[@airhead1]

John Doe u

[@unhook1234]

How did you get a recording of me trying to talk to girls?

[@MerthanE]

Ozymandias Incandenza This.

[@coolnoah8183]

Ozymandias Incandenza actually burst out laughing, thankyou for this

[@rutriggereddrinksomebleach357]

Ozymandias Incandenza LOL

[@dumkkid]

6:06 It says 666 and satan wtf

[@dweller9393]

Alex D funny I could have swore I heard "I'm inraged with Christ"

[@ilonnolan9259]

*6* minutes *6* seconds Or *5* minutes *66* seconds

[@lonelypotato520]

6:02 "Fucks sake"

[@matsekratkievits4889]

I kinda heard "Fuck seven ... Cause it's satan"

[@MachineWashableKatie]

I hate the number 7 too.

[@hardwareful]

jep

[@pkmovies92]

THAT'S WHAT I'VE HEARD TOO ;D

[@firstnamelastname-p7g]

fffff...fucks sade....hail satan

[@minty5221]

I skipped trough the beggining to the end, I'm scared. Please help.

[@coccoborg]

Minty me too, what do we do now?

[@AaronRamm]

I wouldn't be so scared, this is fascinating stuff.

[@silverarcher2248]

as long as they proggram it with Asimov's 3 laws of robotics, we'll be fin

[@evoke2976]

calm down, god doesn't exist.

[@jmcmahon1931]

it's exactly what we're doing with smart phones, captcha and voice commands, we're helping build AI unknowingly with Google.

[@stilltrapped5588]

12:28 is it saying "this is the only thing I can say?"

[@paintspot]

Oh my gosh obviously. Did you just skip to the end and not watch the video? One of the things they trained it on was the repeated phrase "This is the only thing I can say." -Paintspot Infez Wasabi! Like if you agree Reply if you've heard of me

[@want-diversecontent3887]

Paintspot Infez Saw you before.

[@xtrashocking]

Paintspot Infez *I'm not liking, attention brat.*

[@paintspot]

Well good on you for combatting the system! Sorry about that abrasive and attention-begging signature; lost a bet a while ago and now I have to keep this awful signature up. -Paintspot Infez Wasabi! Like if you agree Reply if you've heard of me

[@Kakerate2]

If it was saying that without it being in the training data, that wouldve been really creepy!

[@user-df4zw7yb4v]

4:59 mmmmmnn mmmmmnn mnnnmnmmnm

[@Heliocentricity]

When you feed your neural network some yummy microchips

[@SlimThrull]

This thing will give me nightmares if I listen to certain sections long enough. I wonder if the uncanny valley extends to audio stimuli as much as it does to visual.

[@smurph123452zq]

you should try reading a book to the network so that way it has actual conversation and more information to go off of

[@funny-video-YouTube-channel]

Like a passive-aggressive teenager on team speak. No audible difference :-)

[@davidwuhrer6704]

There is probably a reason why there is no difference. On the internet, nobody knows you are an AI.

[@selflessdeath8119]

Friend stuttered once. Teens start making joke out of it, never let up. New kid comes in, and says (Knowing full well they can speak well) "You guys are fucking retards."

[@geoff3610]

alright at 6:06 it says its first cohesive word and that word is satan i aint fuck with that shit.

[@finn_in_the_bin5263]

geoff3610 you sure cause im pretty sure i heard 'wolfenstein' before that

[@mkultrasoldier]

Kayceesprite Yeah but wolfenstein is a name of a game not a word

[@raengraves5760]

Niall Asher If "Wolfenstein" isn't a word because it is a name, then "Satan" isn't a word either. Your argument is falsely created of moot information.

[@bagle58]

Niall Asher and its still a word

[@ajitihaveamicropenispai9788]

Niall Asher Idiot... ABSOLUTE FUCKING IDIOT

[@Falkon303]

This is awesome. Think you could do a video of feeding an AI dubstep over and over? I simply must know.

[@MaeBlythe]

If you ever need anyone to record hours of repetitive voice-over, I'd happily contribute. Your projects are awesome :)

[@SomethingUnreal]

+Mae Blythe: Thanks for the offer! I'll keep it in mind in case I need that in the future. And I'm glad you like this strange stuff I do =D

[@stumbling]

From 2:50 it sounds like a mixture of Latin and German, which I guess you could say is English! but I mean it flips between the two.

[@billybbob18]

CowLunch English is derived heavily from Latin and German. There might be something to this. It was able to replicate lots of German sounding audio from spoken English. Weird.

[@Yal_Rathol]

billybbob18 not really. German and English are heavily related, to the degree that many of their words can be (and often are) interchangeably used in both languages. generating german sounds from english is as strange as generating green from yellow through trial and error.

[@AmyS3uk]

actually english is a descendant of german. or saxony to be a bit more precise. way back in the days there was no english language but it derived from the saxonian invaders/settlers that settled down on the at that time empty english island/s, so in fact english/brittish people are all kind of german descendants and so are americans if you follow the lineage... there you go....

[@Yal_Rathol]

AmyS3uk germanic, not german, and anglo, who were a different group entirely and settled the british isles first. this is why the term "anglo-saxon" exists and where king arthur comes from, the anglo king driving out the saxon invaders. modern english is a mixture of a few languages, latin and it's derivatives for one, germanic roots, unique roots in the anglos, plus the languages of the celts and danish peoples, but it's base is in germanic culture (who's direct descendant is germany, much like how rome's direct descendant is italy) and anglo culture. this split origin is supposedly why we have different names for animals and the meat that comes off them. "cow" is anglo, "beef" is saxon. this absurd development is why english is so messy and why we've been slowly fixing it over the centuries.

[@billybbob18]

Yal Rathol Agreed, Spanish if my favorite language in terms of logical structure aside from "feminine" and "masculine" pretexts.

[@sett_l5341]

sounds like a discord server when everyone is spamming text to speech

[@LiftPizzas]

Interesting to think that this is generating "speech" which is then being run through a speech recognition algorithm (another neural network?) to generate the auto-captions, which thinks it is recognizing words. The two neural networks are talking to each other (and planning our demise, no doubt.)

[@MaxLohMusic]

DAMN I'M GONNA HAVE NIGHTMARES! p.s. Use the Wavenet algorithm! And a violin! See if you can outclass state of the art sample libraries and put them out of business ushering in the new fun toy for musical composers to get a realistic violin sound given the notes as input.

[@Hawkpathh]

Now that would be real interesting. Those libraries sure aren't cheap... It'd be cool to see them (legally) available for newer composers, along with being synthesized instead of sampled.

[@SomethingUnreal]

I have tried using an implementation of WaveNet, on voice samples, but the most I've got out of it is what sounds like an angry swarm of wasps, i.e. something was very wrong. I've since discovered an implementation made by someone else, so with any luck, that one works better. =)

[@giancarloiannizzotto2980]

I experienced similar issues. Several implementations of the same algorithms, mostly bugged. I recently tried wavenet-tensorflow. It took a lot of time and eventually produced mostly noise. Probably should train more.

[@ryla22]

the program broke down like I am I'm starting to worry that I'm just a program like this becoming self aware. and the scarry thing is, it might actually be happening I have no proof that I actually exist because there is no way to prove it.

[@cleitonfelipe2092]

"This is the only consistent thing I can say" kinda scarry

[@surajtomar4870]

TUN-TUN TUN TUN-TUN. TUN-TUN TUN TUN-TUN

[@andeoo]

the technology for a robotic rebellion is not far from our reach.

[@insideblankoutside]

Adnidor ummm.... It's very far.

[@andeoo]

shoot...

[@MastaSmack]

Not really, he most likely rigged it to say that...

[@chrisgill261]

This is creepy as hell. Why am I listening to it at 04:30 am?

[@vojja01]

Chris Gill Why am I listening this for 5 minutes and reading this at 4.36? o.O

[@Crroow]

Look up poetryreincarnations here on YT. Its kinda satisfying and creepy and the same time. However, this sound very much as Alexander Graham Bell's voice recording from 1885.

[@Sallpls]

Chris Gill 2:38am here.

[@7XHARDER]

Chris Gill wtf it's 5:30 am here lol

[@ryanwallace8263]

2:28 "Shit on Derek's pants"

[@pizzabunk]

2:31 this is so sticky

[@pyroparagon8945]

Where can I get this software? I want to make my school computer sentient EDIT, NVM, looked at discription

[@pyroparagon8945]

8:15 this is why I think this shit is sentient, it's becoming CONFIDENT, and I learns how to not change pitch by itself

[@xlr8is2fast]

ｔｈｉｓ ｉｓ ｔｈｅ ｏｎｌｙ ｔｈｉｎｇ ｉ ｃａｎ ｓａｙ

[@s4ndwichMakeR]

Reminds me a bit of the HAL reactivation scene in 2010: The Year We Make Contact, which was probably more realistic than the DEactivation scene in 2001 where the tempo and pitch of HAL’s voice just decreases.

[@donniemontoya9300]

3:36 "Then you open the car, a gentleman walks up ... He's got a goat" WHAT THE HELL?!?!

[@raptorguy3493]

Donnie Montoya hahaha

[@alexfleming7614]

Donnie Montoya Close Captioning is translating it as "10's as having such a hard time thrusting"

[@peepersmcjeepers]

OH MY GOD

[@crippled82]

Thiago Braga evp 2.0

[@arthinox3317]

sounds like "by him, he's just got a goat"

[@fulviomaiorani6568]

I don' t know about you but from my perspective this is really creepy..... everytime i hear that thing getting close to a word my brain freaks out xD Is uncanny valley a thing even in sound now?

[@opsoc777]

Please come back, your channel is one of the most entertaining channels I'm subscribed to.

[@SomethingUnreal]

I'm still here, though (and thanks, I'm glad you like it!).

[@aransmeallie5344]

The Kings Speech: Android edition

[@ruthlessandevil]

Thats amazing, a computer that speaks perfect Danish.

[@NicoPlatter]

The first two minutes sound like Rick from Rick and Morty trying to speak but having no brain

[@NaughtyKlaus]

Hook this shit up to Cleverbot, and give it a voice it always needed!

[@hecko-yes]

So, WaveNet?

[@zxxNikoxxz]

fascinating and frightening

[@theguyinthefunnyhat]

"This is the only thing I can say." That is some scary stuff. Then again, it was only a matter of time before AI was able to do this. Still, not anything we should be that scared of. The machine might be able to synthesize voice, but it has no grasp of context and therefore cannot speak with intent or purpose. I can't wait to see a machine learn to speak and then learn context only to end up spouting memes 24/7. Hahaha

[@i2pjd6hRw5P]

TGITFH apparently you didn't watch the video. The program said that because that's what its input was.

[@theguyinthefunnyhat]

I watched the video all the way through. You didn't read my comment properly. " The machine might be able to synthesize voice, but it has no grasp of context and therefore cannot speak with intent or purpose." covers exactly what you are saying. I was fully aware the phrase was input. What I meant was the surprising accuracy of the dictation and flection expressed from the program.

[@lollergsize763]

TGITFH The goosebumps

[@AlbySilly]

8:33 It likes "Fun"

[@sargentdogeearrape7404]

Someone else said that at 6:06 you hear "six satan"

[@edwardlau892]

It sounds like a mix of "four" and "one".

[@xiphosura413]

That's probably why it likes saying it, it almost fits both "four" and "One"

[@elliottbarth]

Next phase you should teach, *I'm sorry, Dave. I'm afraid I can't do that*

[@JeroenManders]

This is even more fun when watched with automatically generated closed captions!

[@robin-autumn2441]

can you try feeding it an audiobook as training data?

[@beaniepollard8290]

Grayson Marsh that's a good idea

[@cafbud]

Holly Boreham it is

[@Republic3D]

Google DeepMind have already created near perfect synthetic speech. Interesting stuff.

[@victorwade2343]

this is amazing! truly humbling to see a computer learn to speak from absolutely nothing to trying to babble like a baby. I think we can learn a lot about how the human brain learns to speak languages with this kind of data processing!

[@JloBroOFFICIAL]

sounds like animal crossing

[@noname-bu1ux]

James Logan it totally does

[@pepinyostep3592]

James Logan 😂😂😂😂

[@mag51284]

It's just sped up English in the English copies, at least the first game was.

[@MultiSciGeek]

This is really interesting. Where can I learn exactly how this works.

[@sm0na]

neuralnetworksanddeeplearning.com/

[@diegoantoniorosariopalomin4977]

MultiSciGeek check the Siraj raval KZbin channel , he has just started an intro to deep learning to non programmers

[@tyraelarchangel183]

Neural networks, you should also start with Genetic Algorithms prior, since a decent amount of neural networks rely on some principal of the genetic algorithm. (GA's are basically what you'd expect, an application of "genetics" with respect to computing. Many iterations of "agents" are developed, mutated, bred, and "fight" against each other, and each time there is a "survivor" pool in which they repeat the process. Normally the GA is used to mutate and improve the Neural Network. This is not always the case, obviously, but in general when learning about GA's and Neural Networks they are taught in a linked form where GA's are taught first. Sadly a lot of it is simple once you understand how it works, still fascinating, but I've always thought that most of these are no better than a glorified searching algorithm (there are some exceptions of course), but in general I feel for now the limit has been reached with neural networks until we get some fresher creativity or better algorithms into it. I'm in no way undermining neural networks either, they're fantastic tools and great for an incredible number of things, and perhaps one day they'll be used as a foundation for even better forms of AI. But as of now, just as it was in the past, they really are just glorified searching algorithms. You train them on input data, they adjust interior weights, and through that they can be used to evaluate other types of similar data (such as the neural network that can determine a subject of a picture). Incredible, fascinating, and useful, but it's still not "intelligent" in the way that most people see it through rose colored glasses.

[@diegoantoniorosariopalomin4977]

Tyrael Archangel i don't think you know how important neural networks research is now , name the deep learning applications that amaze you the most

[@tyraelarchangel183]

Neural network research, and most research in computer science (and arguably any field) is important. If you're trying to start an argument you won't find one with me. I never once said I undervalued neural networks, or the research that will help them improve even further. My main issue is that when most people see this video, they think it's a computer learning to speak, especially with fancy words/concepts they may not fully understand. When in fact it's a computer algorithm, that takes user input (this persons voice) and adjusts weights to "match" that input, or otherwise produce a similar output. And in essence, that is how MANY (not all) NN's work. A great example is the picture matching NN. It took input from thousands and thousands of pictures to "train" it. For training either the answers were given (i.e. this is a picture of a cat), or some other form of correction was used. And from there it adjusts internal weights which can represent any number of variables, or interlinked variables. Once trained, this NN was then able to be fed a DIFFERENT picture, and if the weights were set well enough, and the picture matches enough training data, it could then isolate certain subjects and determine that the picture was of a cherry, or a monkey, or what have you. Again, it's fascinating, and amazing. But most people think the computer is learning in some incredible human sort of way, and while this is true in essence, it's no where near the level that most people are imagining. I'd consider myself a realist and just hate when research is over glorified, as that leads to over hype, and when the true results are seen it dissuades people from further investing in the research, which is horrible, as such research is incredibly important.

[@authentic6825]

I've never been more thankful for automated captions

[@Mikeological]

FOOOOOOOUR

[@sashperilla]

Mikeology Robot Golf Simulator 2017

[@XSpamDragonX]

Anyone else getting a more German vibe from the generated voice?

[@alphakretin2387]

Blahblah It's taytweets all over again

[@busTedOaS]

no, german here. sounds like english digits mostly.

[@theamarsh4715]

I was definitely getting a German vibe near the start - some of the sounds it created were much more common in German than in English.

[@Yal_Rathol]

explained this to someone else already, but let's do it again. german and english are heavily related, to the degree that many words from each language can be swapped with no issue in grammar or pronounciation. so it's not strange that it generates english or german sounding gibberish when it's trying to replicate english through trial and error.

[@mrmaniac3]

VeryUnfriendlySpoon 9:40 I can testify

[@ZERR0R]

These is something I like about this. I've been looking for this kind of speech for a few years, but couldn't find anything good enough, but this is almost exactly what I wanted. Meaningless generated speech. I don't know why, but it's... I just... It's calming? I don't know how to explain. But it's... interesting? A program trying to speak. It's not doing it wrong. It's just thinking in other way and speaks in other way. It's not its fault, it's yours. It can talk, it's you who can't understand it. A glitched voice, almost eligible, but not yet there... the uncanny valley manifested in sound. Almost human. Almost speech. So, who is it? The program who can't speak? Or you who can't comprehend it?

[@GretuskaLTU]

I swear I heard it say "Satan" and "no survivors" in the number part...

[@caller145]

LoneMaple I heard it saying "stop crying" at some point xD

[@GretuskaLTU]

We should really stop watching these in the middle of the night haha xD

[@davislast7891]

3:42 When computers start performing ancient satanic rituals