DaveClone

[EEVblog]

I'm experimenting with various AI voice cloning tools.
Just tried ElevenLabs and it's, well, not me at all.
No surprise given the limit of 10MB of source learning material and the instant nature of it.
FAIL



https://www.eevblog.com/forum/chatgptai/daveclone/?action=dlattach;attach=1777136

[MK14]

My understanding (from memory and could be wrong), is that a person has difficulties, properly hearing (comparing) their own/real voice from recordings.  Because, unlike when other people speak, you hear your voice more directly and more via direct vibrations and things.

I.e. You are NOT (contrary to a persons opinion), necessarily the best independent judge of how real it sounds.

To me, (although it does sound a fair bit off), it is a (somewhat vaguely) reasonable representation of your voice (via the supplied MP3 file).  But by no means, anywhere near perfect.

Also, I suspect opinions (perceptions) on how good it is (or not), will vary a fair bit, from person to person.

N.B. I agree, it is not especially brilliant (the MP3), and fairly easy to tell it is not you.

EDIT: Here is an article, which attempts to explain the differences, between hearing your own voice, while speaking and a recording.

https://www.wonderopolis.org/wonder/why-does-my-voice-sound-different-on-a-recording

[RoGeorge]

Very few will notice the differences if you overlap that sample mp3 to one of your videos.
Not a perfect imitation, but very close. 

[schmitt trigger]

The genie is out of the bottle.
It is a matter of time where voice clones become indistinguishable, not only in the pitch, tone and inflection, but also in the unique combination of mannerisms that everyone has.

[rstofer]

I'm experimenting with various AI voice cloning tools.
Just tried ElevenLabs and it's, well, not me at all.

From a distance (say California), it's close enough to call it 'Dave'.

Part of the reason that I accept it as 'Dave' is that I don't hear audio from videos on a frequent basis.  The subtle differences are probably lost on me.

It's  'Dave' for all practical purposes.

[PA0PBZ]

It's missing a bit of variation in tone and speed but yeah, it's Dave.

[ebastler]

I would also say it's close enough. The difference from the "real deal" might stem from the text style as much as from the voice itself.

(The sample text is probably from a textbook or such, very matter-of-fact and not "no script, no fear, all opinion". Could you let the AI voice read a short piece of transcript from a typical EEVblog or EEVblab video?)

[MK14]

I would also say it's close enough. The difference from the "real deal" might stem from the text style as much as from the voice itself.

(The sample text is probably from a textbook or such, very matter-of-fact and not "no script, no fear, all opinion". Could you let the AI voice read a short piece of transcript from a typical EEVblog or EEVblab video?)

That's a very good point, and idea.  With interesting results, if carried out.

[Kim Christensen]

It does sound a bit like Dave... needs some more phrases like "Flappin' in the breeze", "brilliant", "there you go", etc for more realism. 

[gamalot]

For me, a non-native English speaker, the voice is very close to Dave's video, but the accent and tone of the speech are a little different, it sounds too formal or serious.

[SiliconWizard]

It's a lot more lifeless than the real deal, but the voice tone sure sounds very similar.

[Zero999]

I say it doesn't sound like Dave at all. The accent is much weaker. It sounds more like standard British English, than Dave's Australian accent.

[golden_labels]

I can’t tell the difference in the voice itself. But the way of speaking is wrong. Hearing this, I imagine Dave being coerced into giving a lecture to an empty room, while he himself is falling asleep. Give the real Dave a white, empty sheet of paper to describe and he will sound more enthusiastic.

I am also used to hearing Dave at 1.5×. This sounds sooo sloooooow at the default speed.

[EEVblog]

My understanding (from memory and could be wrong), is that a person has difficulties, properly hearing (comparing) their own/real voice from recordings.  Because, unlike when other people speak, you hear your voice more directly and more via direct vibrations and things.
I.e. You are NOT (contrary to a persons opinion), necessarily the best independent judge of how real it sounds.

I have literally thousands of hours of training hearing my own voice through speakers exactly as you and everyone else has. Remember, when I edit my videos I'm listening through speakers, I'm not hearing my voice internally.

[EEVblog]

I say it doesn't sound like Dave at all. The accent is much weaker. It sounds more like standard British English, than Dave's Australian accent.

That's exactly how I hear it, it's clearly british.

[Kim Christensen]

That's exactly how I hear it, it's clearly british.

To us poorly educated North Americans; Brit, Aussie, NZ, ya'll sound the same! 

[EEVblog]

Ok, try this:

[Kim Christensen]

There is definitely a difference, but it's still an impressive imitation. Your real voice is more animated and dynamic and the accent is a bit different, but I'm sure some people could be fooled by the AI voice.

[EEVblog]

There is definitely a difference, but it's still an impressive imitation. Your real voice is more animated and dynamic and the accent is a bit different, but I'm sure some people could be fooled by the AI voice.

Maybe some, but it's so not an aussie accent it's not even close in that regard.

[EEVblog]

I wonder if there's any public figure or private person whose sense of identity can survive the coming wave of too-cheap-to-meter targeted advertising and entertainment.

No person with anything more than a few minutes of public audio of their voice out there is safe.
Wait until I try the higher end ones that build word lists and take days/weeks to process and build a voice profile, I'm betting it'll be chalk and cheese against this simplistic one.
This one literally look zero processing time. I uploaded the source files and that voice you heard was generated a minute later.

[golden_labels]

Dave: in this video having to take a breath makes it sound like a human.

karpouzi9: for the particular situation you mentioned there is a working, well tested and widely deployed solution. Created over 40 years ago. It’s called a public-key signature.

In general case, I see it from a different perspective. Finally the majority can no longer turn a blind eye to problems, which were present for a long time, but were either understood by or affecting minorities or people deemed not worth caring about.

[Circlotron]

No person with anything more than a few minutes of public audio of their voice out there is safe.

Pity Bob Hawke isn't still Australian Prime Minister then. Comedians and others had a great time with his voice back in the day.

https://youtu.be/xz9BmAXW_Gs

[David Aurora]

Yeah, nah. I'll give it points for timbre- for a split second I was impressed on a word or two (FFT analysis of your voice vs the fake voice could be pretty interesting), but the accent difference was hilarious.

Still, this time next year it'll probably be perfect

[EEVblog]

Pity Bob Hawke isn't still Australian Prime Minister then. Comedians and others had a great time with his voice back in the day.

Totally iconic!

[golden_labels]

It’s worth remembering, that person’s voice should not be considered secret from security perspective. I am mentioning this, because the general population believes information not offered to wider public audience is secret. Just like other similar instances, one’s voice is both both processed by many parties and easily obtainable in active attacks (a phone call, for example).

Regarding fakes, a recommended movie: “The Congress” (2013).