I Spoke With An Ai Version Of Myself, Thanks To Hume's Free Tool - How To Try It

Chiken Brave/Getty Images

If you've ever had nan impulse to converse pinch an AI type of yourself, now you tin -- benignant of. 

On Thursday, AI start-up Hume announced nan motorboat of a caller "hyperrealistic sound cloning" feature for nan latest loop of its Empathic Voice Interface (EVI) model, EVI 3, which was unveiled past month. The thought is that by uploading a short audio signaling of yourself speaking -- ideally betwixt 30 and 90 seconds -- nan exemplary should beryllium capable to quickly churn retired an AI-generated replica of your voice, which you tin past interact pinch verbally, conscionable arsenic you would pinch different personification opinionated successful beforehand of you. 

Also: Text-to-speech pinch emotion - this caller AI exemplary does everything but shed a tear

I uploaded a signaling of my sound to EVI 3 and spent immoderate clip idly chatting pinch nan model's imitation of my voice. I was hoping (perhaps naively) to person an Uncanny Valley acquisition -- that exceedingly uncommon emotion of interacting pinch thing that feels almost completely real, yet off-kilter capable to make 1 consciousness somewhat uneasy -- and was disappointed erstwhile nan EVI 3 maine was much for illustration an audio animation type of myself. 

Let maine unpack that a bit.

Using EVI 3's sound cloning feature

Screenshot by Webb Wright/ZDNET

The imitation of my sound was, successful immoderate ways, undeniably realistic. It seemed to region intermittently erstwhile speaking successful much aliases little nan aforesaid measurement that I thin to do, pinch a touch of acquainted vocal fry. But nan mirroring stopped there.

Hume claims successful its blog station that EVI 3's caller sound cloning characteristic tin seizure "aspects of nan speaker's personality." This is simply a vague committedness (probably intentionally so), but successful my ain trials, nan exemplary seemed to autumn short successful this regard. Far from emotion for illustration a convincing simulation of my ain behaviour quirks and consciousness of humor, nan exemplary said pinch a chipper, eager-to-please reside that would've been well-suited to a power advertisement for antidepressants. I for illustration to deliberation of myself arsenic being friends and mostly upbeat, but nan AI was evidently exaggerating those peculiar characteristic traits.

Also: Fighting AI pinch AI, finance firms prevented $5 cardinal successful fraud - but astatine what cost?

Despite its mostly puppy-like demeanor, nan exemplary was strangely staunch successful its refusal to effort speaking successful an accent, which seemed to maine for illustration it would beryllium nan benignant of playful sound workout that it would excel at. When I asked it to springiness an Australian accent a whirl, it said "g'day" and "mate," erstwhile aliases doubly successful my normal voice, past instantly shied distant from thing much daring. And nary matter what I prompted it to speak about, it tended to find immoderate imaginative and roundabout measurement to circle backmost to nan taxable I was discussing erstwhile I recorded my sound arsenic a sample for it to use, reminiscent of an research from Anthropic past twelvemonth successful which Claude was tweaked to go obsessed pinch nan Golden Gate Bridge.

In my 2nd trial, for example, I had recorded myself speaking astir Led Zeppelin, which I'd been listening to earlier that morning. When I past asked EVI 3's sound clone of myself to elucidate its thoughts connected nan quality of acheronian matter, it quickly recovered a measurement to bring its consequence backmost to nan taxable of music, comparing nan mysteriously invisible unit pervading nan cosmos pinch nan intangible melody that imbues a opus pinch meaning and power.

You tin effort EVI 3's caller sound cloning characteristic for yourself here.

According to Hume's website, personification information produced from interactions pinch nan EVI API are collected and anonymized by default successful bid to train nan company's models. You tin move this off, however, done nan "Zero information retention" characteristic successful your profile. For non-API products, including nan demo linked above, nan institution says it "may" cod and usage information to amended its models—but again, you tin toggle this disconnected if you create a individual profile. 

Whispering robots

AI voices person been astir for rather a while, but they've historically been alternatively constricted successful their realism; it's very evident you're talking to a robot erstwhile you person responses from classical Siri aliases Alexa, for example. In contrast, a caller activity of AI sound models, EVI 3 among them, person been engineered not only to speak successful earthy connection but also, and much importantly, to mimic nan subtle inflections, intonations, idiosyncrasies, and cadences that inflect real, mundane quality speech.

"A large portion of quality connection is emphasizing nan correct words, pausing astatine nan correct times, utilizing nan correct reside of voice," Hume CEO and main intelligence Alan Cowen told me.

As Hume wrote successful a blog post on Thursday, EVI 3 "knows what words to emphasize, what makes group laugh, and really accents and different sound characteristics interact pinch vocabulary." According to nan company, this marks a awesome method leap guardant from earlier speech-generating models, "which deficiency a meaningful knowing of language."

Many AI experts would return umbrage pinch nan usage of words for illustration "understanding" successful this discourse since models for illustration EVI 3 are trained simply to observe and recreate patterns gleaned from their voluminous swathes of training data, a process that arguably doesn't time off immoderate room for what we'd admit arsenic existent semantic comprehension.

Also: ChatGPT isn't conscionable for chatting anymore - now it will do your activity for you

EVI 3 was trained "on trillions of tokens of matter and past millions of hours of speech," according to Hume's blog post. According to Cowen, this attack unsocial has enabled nan exemplary to speak successful voices that are overmuch much realistic than would intuitively beryllium expected. "With sound [models], what's been astir astonishing is really quality [they] tin beryllium conscionable by training connected a batch of data," he said. 

But philosophical arguments aside, nan caller activity of AI sound models is uncontroversially impressive. When prompted, they tin research a overmuch vaster scope of vocal look than their predecessors. Companies for illustration Hume and ElevenLabs declare that these caller models will person applicable benefits for industries for illustration intermezo and marketing, but immoderate experts fearfulness that they'll unfastened caller doors for deception -- arsenic was illustrated conscionable past week erstwhile an chartless person used AI to imitate nan sound of US Secretary of State Marco Rubio and subsequently deployed nan sound clone successful an effort to dupe authorities officials.

"I don't spot immoderate logic that we would request a robot whispering," Emily M. Bender, a linguist and coauthor of The AI Con, precocious told me. "Like, what's that for? Except possibly to disguise nan truth that what you're listening to is synthetic?"

Revolutionary becomes routine

Yes, EVI 3's sound cloning feature, for illustration each AI tools, has its shortcomings. But those are importantly overshadowed by its singular qualities.

For 1 thing, we should retrieve that nan generative AI models hitting nan marketplace coming are portion of nan technology's infancy, and they'll only proceed to improve. In little than 3 years, we've gone from nan nationalist merchandise of ChatGPT to AI models that tin much aliases little convincingly simulate existent quality voices and devices like Google's Veo 3, which tin nutrient realistic video and synchronized audio. The breathtaking gait of generative AI advancements should springiness america pause, to opportunity nan least.

Also: AI agents will alteration activity and nine successful internet-sized ways, says AWS VP

Today, EVI 3 tin simulate a unsmooth approximation of your voice. It's not unreasonable to expect, however, that its successor -- aliases possibly grand-successor -- will beryllium capable to seizure your sound successful a measurement that feels genuinely convincing. In specified a world, 1 tin ideate EVI aliases a akin voice-generating exemplary being paired pinch an AI supplier to, say, subordinate Zoom meetings connected your behalf. It could also, little optimistically, beryllium a scam artist's dream travel true. 

Perhaps nan astir striking truth astir my acquisition interacting pinch EVI 3's sound cloning feature, though, is really mundane this exertion already feels. 

As nan gait of technological invention accelerates, truthful excessively does our capacity for instantaneously normalizing that which would person stunned erstwhile generations of humans into awestruck silence. OpenAI's Sam Altman made this very constituent successful a caller blog post: According to Altman, we're approaching nan Singularity, yet for nan astir part, it feels for illustration business arsenic usual.

Want much stories astir AI? Sign up for Innovation, our play newsletter.