It's fun but as I tested it I realized how this is pretty much the modern equivalent of a Facebook quiz that asks you the name of your first pet, first car and mother's maiden name.
If they recorded any of that they likely have enough to clone my voice somewhat faithfully.
Congratulations on labelling my French Canadian accent as French though, I'll have to work on my pronounciation more to fool the AI.
I have a similar experience, but my impression is that Parisians have little patience for accents in general. People elsewhere in France where I've visited (Aix-en-provence, Lyon, Toulouse) had no trouble chatting with me.
Also want to note that I am not passing judgement, I am sure I might become a bit brusque with tourists if Montreal had the same volume.
My wife was a French major in college and spent a year in Aix in the mid 80s. She didn't really enjoy the experience -- although her grammar and vocab were fine, she had enough of an accent that people treated her as an outsider.
On the way out of the country she stopped in Paris for a few days to take it in. She went into some shop and the proprietor asked my wife if she needed help. My wife replied asked about some clothes. The proprietor insulted my wife over her southern accent -- French southern, not US southern, and it made my wife's day. Someone in France finally thought she was french.
Does it regionalize at all? There are a few comments about Portugal Portuguese vs Brazil Portuguese so I think it only tries to find your first language, not actually pinpoint the underlying accent.
Ya it doesn't seem to go down into dialect, its the broad scope, it nailed my wifes spanish accent even though i really can't hear it at all and she talks very american
It would be nice if they were clear they wouldn't keep the sample on the page. They do have a privacy policy on their main site www.boldvoice.com/privacy
It didn't guess for me other than to say I was a native speaker.
Hi, founder here! We're a YC-funded education app with over 1 million downloads, currently #7 on the Education chart of the App Store. The goal of the accent oracle tool is to help more non-native English speakers find out about BoldVoice. While it's important as consumers to be aware of data privacy with AI tools in general, I want to set the record straight that that's not the point of what we're doing here.
But great to hear that we got your accent right :)
Hey! Just to be clear I really didn't want to imply that you were trying to clone my voice. I think your product is very interesting and this demo is a nice showcase of your tech.
It was meant more as a reflection on how, due to how fast threat models evolve, even the seemingly innocuous act of recording 2 sentences on a random website can now be used to break into my bank account.
I tried the app, and it's great. As a near-native French speaker I would be interested in working out my French pronunciation, hopefully you will go multi-language soon.
Did you know that something like 80% of the websites you visit are tracked by a handful of companies which are all effectively sharing that with many thousands of ad and data broker companies which can easily narrow you down? It's not tinfoil hat land to knkw that nearly every site you visit has cross-site tracking.
This was a big confidence booster for me as when I first started learning English, people would complement me on how well I spoke English, but I took that as my accent was still detectable. It's only been in the past 5 years that people assumed I was American and made no comment on my English at all, until I disclosed that English was my second language. It's usually certain words that give me trouble, like "cupboard" or "chef". The AI detected my accent as a mixture of German and English. When I tried to exaggerate my accent, it correctly detected Thai.
If you learned English after 16. You probably still have an accent. Native speakers are really, really, really good at detecting it. They probably know as soon as you say "Hi".
If you tune up your detector that high it’s very likely you’ll get false positives. I’ve met Americans with fairly strong “accents” that aren’t necessarily a dialect either, just a different way of speaking. It could also be a mix with parents who are non-native. American English is vast, and very heterogeneous.
If we’re talking about specific parts like a regional dialect then I would agree, those are tricky to acquire later, at least to those undetectable levels. They can be extremely specific.
Yep big difference in accents of my cousins who moved here when they were 9 and another when they were 18. Now they are in their mid forties and you can still tell who moved when based on their accent. Its impossible to change your accent after late teens.
there are people who are better than other people at blending in their accents, even in more difficult languages than English (for accent coding), perhaps they are just very good at that.
those words are your Shibboleths, words that give your origin away.
When I was in Germany, friendly people used to compliment me on my language skill saying "your German is good!". To which I would reciprocate: "thanks, yours too!"
My ex-wife whose native language is Spanish worked hard to eliminate her accent because she got tired of people calling her accent “cute.” Her shibboleths were anything with a schwa. The whole concept of schwas offends her sense of vocalic purity.
I knew my accent was strong, but I didn't expect to get 100% Portuguese, which is strange since Portuguese from Portugal sounds more like Eastern Europe, and Portuguese Brazil is more like Spanish. Maybe it considers both accents to be Portuguese?
A fun fact: When using Whisper by OpenAI, there seems to be a ~1% chance that all my text, which was spoken entirely in English, is automatically transcribed and translated into pt-BR without any prompting. It happens more often when I am not paying too much attention to pronunciation.
The weird thing is that all the words were transcribed correctly (beyond being entirely in a different language)
> Portuguese from Portugal sounds more like Eastern Europe, and Portuguese Brazil is more like Spanish.
Surely you mean the opposite? Portugal is literally next to Spain and both languages have coexisted since they were both born following Rome's fall. Both Galician and Portugal's Portugese are likely similar to each other and closer to Spanish than Brazil's Portuguese
You are right regarding Galician, but Galician isn't Spanish, rather original Portuguese, where some words changed throughout the centuries.
The way we both speak is rather different than the native Spanish speakers, that never have been exposed to native Portuguese/Galician speakers.
In fact, even native Catalan speakers have easier time understanding Galician/Portuguese speakers, than Spanish speakers do, probably due to the French roots in Galician/Portuguese carried by the crusades involved in the founding of both regions and naturally influeced the language evolution.
One cool fact is that Galician sounds closer to Brazilian Portuguese. It also has much of the same vocabulary (with the notable exception of the days of the week which it borrows from the other Romance languages while Portuguese has its boring days of the week)
Presumably each training speech sample is labelled with native language. For Portuguese there would be two distinct clusters: Portugal and Brazil. If your speech is in either cluster, it would just tell you that your native language is Portuguese without being any more specific. Sure, it's a missed opportunity but it doesn't distinguish Jamaicans from Australians either.
I presume there's enough difference between English spoken by Portuguese and English spoken by Russians for those also to be distinct clusters.
The homepage sort of implies that "having an accent" is something only non native speakers do? Like an accent only comes from your exotic mother tongue. Kinda weird. It told me I'm a native speaker, and I am a non American native speaker so... good I guess?
I don't follow. Why is the data they are getting from this better than the billions of hours of captioned voice data available from youtube/tiktok/instagram/whatever?
I don't mean to suggest that this particular toy is much use. I just mean you should not give random internet games standardised samples of your voice, for this reason.
It's a standardised sample, already correlated to text, close to the microphone, for one thing. You're just making it easier for them.
I mean I suppose you can use "like and subscribe", "without further ado", and "let's get started" as standardised samples if you want to catch a youtuber.
But AFAIK my voice isn't on the internet anywhere. Quite a lot of people are not.
There's a number of ways this information can be connected back, with varying precision, to the person who recorded it.
And we should have learned from the Cambridge Analytica scandal that data is used in ways we do not expect. For example, what if you don't care to reproduce someone's voice, but you do care to extract age/gender/racial background/sexual orientation from it?
What I mean is that there are still billions of people whose voices are not on the internet.
I'm more than half a century old, an internet geek since a few years before the conventional "dawn of the web" and AFAIK there is no recording of my voice on the internet.
Added to which, controlled samples like this with a good range of syllables will always be more helpful, won't they?
Or teach people not to trust that you said something just because something sounds like you. Use actual authentication instead of implied. Same for photos, videos
This is perfect for CEO scams in most American companies.
Many (large and small) American companies (and other nationalities as well, sure) a top down management approach is the norm. I.e. "CEO" (or "your manager" / "person in power") says something and you jump and do it without asking any questions because you fear you'll be fired otherwise (or have other repercussions).
In such an environment, imagine the CEO / person in power giving the best sample ever to the crooks, such that they can clone your voice almost perfectly. Now, of course, CEOs are likely to be recorded in various events anyway but some others are less likely, say the CFO.
Then order some lowly finance drone to wire a billion bucks to your account (well, maybe a bit less, and make sure to use someone else's account, seven levels of money mules and 17 different crypto currencies with mixers etc. before cashing out) with your faked voice.
We caught a CEO scam that was pretty good but noticeable recently. They had cloned his voice.
Isn’t it desirable to weed out organizations with such fragile procedures…?
It’s like how those ransomware thieves incentivize all the critical computer systems in the world to remain air gapped, which seems like an overall net positive.
In a sense I agree with you. However, really great organizations have weak links. It only needs one unfortunately. I personally don't want to be out of job because of one weak link.
Sort of to your point, we do have training (which I find obnoxiously dumb, but many seem to find it great - I just let the video run in the background and answer the questions without actually watching a single second of it) around this sort of thing and we have phishing tests that are super easy to figure out (the email headers literally tell you it's a phishing test) but various people post on internal channels "Is this a scam? I'm not sure, please help!" and not all of them are non-technical people at all.
Above a certain size of company there just are gonna be some weak links in just the wrong place(s) randomly even with the best procedures unfortunately.
Now all they need to do is somehow work out who you are from only your IP - no email, name, ___location or anything - then simply get a voice cloning model to work perfectly from this small sample, then either somehow hack all the other information needed to get into your bank account or chase down your family to get them to send them crypto and they've got you dead to rights. Simple as that, which is why I also never take phone calls, pay for anything with a credit card or go outside.
I have often wondered how much of someone's actual voice leaks through into their impersonations in a way that can be detected.
I just saw an incredible Facebook reel of a voice actor, Shelby Young, saying the same thing in a striking range of theatrical voices, and I still wonder. How much of her true vocal fingerprint is unavoidably there?
(As a fan of old movies, "Vintage" was particularly impressive to me -- she is impersonating not just voices but also the choice of tonality those actors made in light of recording technology)
I'm usually able to identify most of the big VAs regardless of the disguise they're putting into their voice. Dan Castellenta, Jon Lovitz, Maurice LaMarche, Rob Paulsen, Jim Cummings all come to mind as ones I've correctly identified in various productions. The list is larger I'm sure.
An AI might be able to gain a higher success rate than me.
You enter your First Name, Last Name, Gender, Date of Birth, Pet's Name and Mother's Maiden Name and press the button to find out what your Mr T Name is...
The app this is advertising helps non-native speakers with their accent, I assume to sound more American. This is a great goal, and I'm sure there are a lot of people who would be willing to pay the $200-$300 yearly subscription cost. Apparently the AI part is not even the main function of the app, that's what the extra $100 are paying for[1].
I would be interested in an AI-only product that would help me learn to passably immitate various English accents, like Australian, Irish and so forth, for fun. I know that ChatGPT Voice can do accents pretty well, I've been wondering if it would also be able to help me with mine, but I haven't tried it seriously.
I could absolutely see people be willing to pay for this. I am from the Midwest in the United States and I happened to be at an airport in some foreign country. Someone else heard me talking and they came up and asked me where I learned to speak English because it was so smooth. They were looking to get lessons to make their English better or at least more smooth. I thought their English was fine and they were a bit disappointed when I mentioned I was from the United States.
It's kind of annoying when services like this provide a free trial that you have to give a credit card number to even try, capitalizing on people forgetting to unsubscribe after trying.
Also, I'm very suspicious when a credit card form is on $site.com rather than $financial-institution.com
My mind is blown right now. My whole life I've been told that my speech is so American and that I don't have a Russian accent (left Russia when I was 4). Lo and behold, this app tells me that my accent is Russian (61%) with English being a distant second (13%).
I tried it and it said English 93% (left same age as you).
Then I did my best Russian accent and on the first time it gave me Hindi/Urdu at 80+%. I tried it a second time rolling my r's a bit more and it settled on Russian at 70%.
I think it's very sensitive to specific tells and I suspect the dataset for Russian accents may not account for all the variations in regional pronunciation and dialects.
I left Russia around the same age and got 100% English. I can easily do a fake accent and get Russian though. Also some other accents like German, French and so on are pretty easy to get too.
It gave this native English speaker "Swedish" with p > 90%. Just confirms the feeling I get every time I go to Sweden that they really do speak better English than me.
It thought my native language was Hindi/Urdu which was amused me if only because whenever I try to do a foreign accent it eventually morphs into a Hindi accent no matter where it started.
Nice suggestion! BoldVoice focuses on helping non-native English speakers to learn the American accent, so we tailored the accent oracle to non-native accents specifically.
Have you thought of expanding to native American English speakers who want to learn a more "standard" accent? There are many Americans who feel that their strong regional accent (think e.g. Chicago, Mississippi, Baltimore, and different black accents) can be career-limiting.
I’m not suggesting you change your product. I’m simply explaining what I personally would be interested in.
That said, if you want to teach people to speak with an “American accent”, you should probably be aware that there is no “American accent”, only regional accents.
Thanks - we're aware! There's also a concept of a general American accent. Since the app is for non-native English speakers, a general American accent is the goal more so than a regional accent. We built the app together with professional accent coaches -- if you're interested in our methodology, you can learn more on boldvoice.com.
Not very good guesses. It had me read twice and I used a high quality mic. It guessed Spanish as my native language, but picked up a bit of Chinese and a bit of English. I am a native born American whose only language is English and a life long Midwesterner. I have a midwestern accent and occasionally some Canadian influences sneaks in (or so people have told me), but Spanish/Chinese? Completely wrong.
As a non-native speaker, it is very very accurate even for Hungarian, which is quite a tiny language. I have sent it to several friends and it "caught" them all.
Whenever time permits, I have a (bad) habit of viewing source code of new website. In this case, I found this on this website: (haven't read the js yet to see what's the true intent but surely a sign of horrible engineering)
My guess is that it's from their SSR framework (ie. remix), which serialized way too many things and sent it to the client. That, and they're using the same feature flag project/config as their main app, because looking at the feature flags it's clearly to do with their main app (ie. AI voice training) rather than this AI voice guesser app.
> "successEnterReferralCodeDuringOnboardingBody": "You've just unlocked 10% off your BoldVoice subscription, thanks to [firstName]'s referral!",
> You sound like a native English speaker. I couldn’t identify any distinct non-native accent.
I am a native English speaker, with an Australian accent. I think it's supposed to identify your non-english-native accent, which you wouldn't have one being Australian.
Probably should be read as "bad actors", who, with enough samples of your voice, could theoretically do shady things like robo-call your mom, pretending to be you.
Interesting, I was hoping it would be more specific than "English" (e.g. "Southern Illinois"), but I'm sure that's just around the corner. It looks like this is an advertisement for a product to "lose" your accent, so as long as you sound like a native English-speaker they're happy.
I tried again using an outrageously bad (probably to the point of offense) Scottish brogue and it pegged it as German.
Would be cool if it could detect area specific accents. I grew up in Kentucky and tried it in a very thick Eastern KY accent and it just said native english speaker. (technically true)
Here's what it sounds like:
https://www.youtube.com/watch?v=bB8vHRH9A6M
It doesn't work - it seemed to catch my accent once I ran it a couple of times on the same browser, but... in the incognito mode, it fails every single time, ranging from French and Swedish to Swahili. :D When I mumbled a bit and lowered my voice, it said 97% English, lol. Maybe its model treats mumbling as the UK, hard to say. When I added a bit of "R," it immediately recognized me as Russian, ignoring everything else. And when I increased my pitch without changing anything else, it started detecting Persian and Spanish. When I used a proxy server pointing to Scandinavia, it started detecting Swedish, lol. Fake as hell. I'm Polish, btw.
First, they are not same thing. If they would be same thing, russians could speak or understand Ukrainian language (they can't).
Second - I don't have russian accent. I don't even speak russian. I could agree that it can be 'slavic accent' but why attribute it to russia, majority of which is non-slavic literally Asia?
The AI couldn't guess my accent correctly which is OK as it's fairly non-standard. However, the onboarding flow needs work. I feel like it took too long to ask too little and it made the wrong assumptions.
The app also crashed towards the end.
NotFoundError: Failed to execute 'removeChild' on 'Node': The node to be removed is not a child of this node.
at https://start.boldvoice.com/build/_shared/chunk-NFFSPFRU.js:1:627
at Ti (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:22278)
at _t (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:23972)
at Xn (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:41320)
at Bf (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:40880)
at hn (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:39936)
at Qo (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:36620)
at pn (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:6:3250)
at Bf (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:40935)
at hn (https://start.boldvoice.com/build/_shared/chunk-WREYPQ4L.js:8:39936)
re:assumptions I realize my experience is outside the norm. But there are "native" english speakers in most countries on Earth. Immigrants and expatriates are an example of one such community.
The app assumes that there's only one kind of "native" speaker i.e. Americans, British folks and Australians. That's not the case. Over 80 countries have native english speakers. Many of them have accents that are outside the American and British norm.
This was less interesting than I was hoping for, because it wasn't specific. It said I don't have a non-native English accent. Great, I already knew that. But I'm curious if it could place my regional accent in the United States. I'm originally from the Southern Midwest which has a distinct accent, but have made a great effort to neutralize my accent and believe I now sound neutrally American (what used to be called Nebraska Newscaster).
Sounds like others tried and had similar results (not identifying Australian or Irish accents).
Yep, as a New Zealander/Australian (spent half my life in both) I was curious what it would give me. Turns out "native english speaker" is all you get. Even if I really put it on thick
I've tried the app (free trial), it's doing a good job at identifying issues in my pronunciation. It can rate my level and highlight the vowel / consonants within words that need improvement. The app looks quite good, but a bit expensive. I'd be very interested if it supports more languages in the future.
I suspect they use the same technique for guessing the accent - detect which sounds are not well pronounced (and they have no interest in distinguishing accents amont native speakers).
When I intentionally spoke in my native accent (which is not something I normally do), it guessed it with 100% confidence, even though it's not very common. Impressive.
When I spoke like I normally do, it wasn't able to get anything on the first try, and on the second attempt it guessed 3 very different accents (e.g. Danish, Persian, ...) with more or less equal confidence. But it didn't guess my native accent at all.
Huh, I always thought I sound almost American. Looks like my accent is untraceable at best.
Speaking normally it identified me as English (seems it just means native and not actually specifically English?). Putting on a Swedish accent it got that too, but if it really picked up subtle details as claimed it should've identified that as English too as I break waaay too often with certain vowels.
I'm surprised it considered my truly awful American native, but it needed two clips to decide that time. And 30% Hindi/Urdu? What?
Well, I am not a native English speaker and it says 77% English (and 6% Chinese, 5% Spanish). I guess I should take that to mean that my English is pretty good.
I haven't tried it, as you get asked for the answers to common security questions, and supply a voice sample. You could lie but many people won't.
My other issue is that it will have been trained on a large number of voice samples and no one will learn how to distinguish different accents by using it, or even by developing it.
An alternative, knowledge-based approach, would work by splitting the speech into phonemes and matching phonemes/sequences of phonemes against known accents or foreign languages, e.g. if a native speaker rhymes "good" and "food", you can reliably tell they're either from Scotland or Ulster. Telling close accents apart is easy with the right phrases, e.g. "fish and chips" (Australian vs. New Zealand), "I saw the White House" (General American vs. Canadian). For non-native speakers, you can use the phoneme inventory of their native language, so if someone has difficulty in pronouncing "th" you can rule out Greek or Spanish (from Spain), and if someone has difficulty in pronouncing "f" they're probably Korean.
Of course, that's a lot of work up front, but you'd learn a lot in the process of developing such a system, and it would be able to explain its decisions to users. And by asking you to repeat standard phrases (like "good food") you would allay security concerns.
The app this is kind of a PR for does phoneme-level analysis, so for all we know, this AI could be doing that as well. See https://youtu.be/j6z2WHAvqEs?t=100
This tool works pretty well, it guessed me right as well as few of my coworkers who are from a different parts of the world and none of us have obvious accents. This is scary good but I'm afraid privacy will be impossible in the future, we'll be analyzed and categorized instantly. The only barrier to completely losing privacy is our own thoughts.
I'm happy with my mix of Urdu + English accent, I got an 80% on Urdu which seems about right. I am impressed by this and now I'd like to hear others and how well it matches their voice. Although, I don't need coaching or anything to remove my accent, it makes me me.
I am from south jersey and close enough to philly to have a similar accent. I have been traveling and had people pinpoint where I was from multiple times.
Its making me wonder if my reading voice is more proper. Or possibly the thing just doesn’t work.
I'm not a native English speaker, but that's what it guessed.
Over the years, starting in my late teens (I'm in my late 30s now), I've put TONS of effort into sounding like a native speaker (moving to the US 10+ years ago has certainly helped).
I'll be impressed when it can tell me I have a California English accent. Surprised it doesn't even discriminate the vast pronunciation differences between British, Oceanic, Caribbean, African and American native English accents.
Complete garbage. Native USA-born (but well travelled) English speaker and it tagged me with Swahili (24%), French (19%), and Persian (19%). I was speaking quietly because it's late here, but that shouldn't have made a difference.
<<Your accent is Persian my friend. I identified your accent based on subtle details in your pronunciation. Listen to your audio, and bask in my predictive abilities.>> --- Wrong, my friend! I'm Brazilian and I speak Portuguese.
Hmmm... it doesn't differentiate between English accents, like UK (actually there are a bunch of sub-accents), Canadian or US (of which there are a bunch of sub-accents.)
It only does non-English accents I guess. That wasn't clear.
Sounds a bit too much to ask. Getting enough labelled training data for each English speaking country (why would you even class the UK as one when England alone has dozens of accents?) in the world is likely a challenge
> Sounds a bit too much to ask. Getting enough labelled training data for each English speaking country (why would you even class the UK as one when England alone has dozens of accents?) in the world is likely a challenge
I am not sure that is the case. There is an AI accent generator that can do different English accents, sort of what I was suggesting to identify:
Interesting that it actually guessed as the neighbouring country. I live close to the border but have never interacted with the people across. Wonder if there might be similarities in tone and inflection.
What I really want is an AI to translate difficult accents into ones more familiar to me because there is lots of content that is just too taxing for me to listen due to the mental load.
Fun to see this in the wild! I'm one of the co-founders of BoldVoice, AMA :)
p.s. I'm Albanian and it guesses either Albanian or English for me, sometimes randomly Spanish.
Conspiracy theories are fun! But we're a language learning app (YC s21). We help non native English speakers to work on their accents so they can speak English confidently. The accent oracle is a marketing tool for us to get in front of new users.
I'm genuinely surprised it got my accent right. Coming from Serbia, I'd never expect to get it right. My first guess was that it's geo-ip based, but I could be wrong.
Really fun is trying to do various accents and seeing what it thinks you sound like, I tried my best South African and it thought Japanese. Guess I need to work on my mimicry
I'm half South African and can do many of the different accents, notably Afrikaans, Cape Town (white and coloured) and black SA, they're all very different. I'm pretty sure most people think of Wickus v.d. Merwe from District 9, but that's not the only one!
I can also speak a bunch of other languages without accent so good luck for any AI to fit me in a box :P
The big question now would be: has anyone used BoldVoice or any other method to convert from 90%+ of accent A to 90%+ to accent B and can switch between those seemlessly?
The first guess was shockingly spot on. Tried it again and it thought I was English. Tried it again and it thought I was Dutch. First guess was right though.
95% german, 5% dutch. Guess I can't stop schnacking on platt to some degree.
Though this is raising the fun question: What makes a german accent in english? Harsh consonants? Is there some wiki or some articles to read up on that?
Holy shit. I grew up in Armenia when I was 8 and been in the US for 22 years, and by all accounts English is my primary language, and this got it spot on, with 84% confidence. Was not expecting that for such a unknown accent.
I am guessing this was not trained on a dataset of people speaking English in various accents, but rather is directly detecting your native language.
It guesses Swedish for me. I'm Norwegian. While they have some similar quirks, like sounding "sing-song-y" to a lot of native English speakers, Swedish and Norwegian English accents are usually quite distinguishable from each other.
Given our (good-natured) neighbor-rivalry I'm of course horribly offended.
If they recorded any of that they likely have enough to clone my voice somewhat faithfully.
Congratulations on labelling my French Canadian accent as French though, I'll have to work on my pronounciation more to fool the AI.