How to defeat telephone voice recognition robots

[Mitch]

In the epic battle I've been waging against my ISP in the last 3 days, I've come across a useful tool for getting through those annoying robot voice recognition things you encounter when phoning a customer "support" service.

These voice recognition devices are usually based on some kind of neural network program - often a simple back-propogation system that loads neural path weightings when a positive result is achieved. Each time they get a successful result leading to the next step in the question/answer schema that results in a correct passthrough to the suitable operator.

Unfortunately, they have a tough time learning all the accents, voice types and levels to uniformly come to correct outcomes.

This is my insight:

As the voice recognition system learns, it usually agravates the customer with every "I'm sorry, I didn't understand that - please hear the options again ...". It therefore follows that the system will probably be learning the voice patterns of very angry customers. This would also result in the human customer "support" operator being patched through to a very agressive, upset and vindictive customer. I can only assume that all those Indian customer support people must think westerners are a complete pack of pigs.

The upside is that if you BEGIN with shouting agressiviely at the voice robot, you will be recognised imediately and routed to the correct operator quickly. Having put on the agro-show deliberately, you can then proceed with serenity onto the matter of discussion with the human operator carrying a sense of victory. God knows - it might also assist international relations.

[Innocent Bystander]

Damn fine idea, Mitch. I shall try that.

[MTGuru]

These voice recognition devices are usually based on some kind of neural network program

Not necessarily neural nets. Non-learning statistical models are probably more common in commercially deployed systems. Don't ask me how I know.

But I heartily endorse shouting into the phone at robots, as a matter of principle.

My approach is to always use my very best, clearly pronounced Gibberish. A few responses of "Glorpmux floopzur" and "Mishpish akkkhhhh" get you to a live agent pretty quickly. Klingon: if you speak it, use it.

Continuing to speak Gibberish to the live agent is optional, but may be just as effective as English in inevitably not getting what you want anyway.

[pipersgrip]

I like your style Mitch. I am going to try that now.

[Nanohedron]

I'll try MTGuru's way. I detest shouting, and besides, it makes me fall into coughing fits. Why compromise my dignity so? There are better ways. If I want to compromise my dignity, Gibberish will suit me to a "T".

[brewerpaul]

I like your style Mitch. I am going to try that now.

Yeah, I like the gibberish idea. You can do that in places where shouting would be inappropriate, eg at work.
I may try some good Yiddish cuss words...

[cowtime]

I don't even have to resort to gibberish... I just talk like I always do.. ....

[Protean]

My favorite experience with one of these was some years ago.

I was trying to resolve some issue, I cannot even remember what or with whom any more. All I remember is that I was, shall we say, annoyed. And this particular system had a heck of a time understanding me. It would generally take at least two attempts to have the system understand me. So, in frustration, I grumbled, "I hate these f---ing things."

To which the system responds, "I think you said you wish to speak to a customer service agent. Is that right?"

[khl]

To which the system responds, "I think you said you wish to speak to a customer service agent. Is that right?"

I'll probably be laughing for an hour over this one.

[Innocent Bystander]

Wonderful. And, a good example of great programming!

[Mitch]

Hmm - I'm pretty sure the robot I was talking to was a learning algorythm - I deduce this by the fact that it responded so well to the "agro-man" show. neural learning systems are quite simple and cheap - most voice doodads use this - for instance windows voice-command, in your mobile or bluetooth - in fact any system that requires a training session to get it all started. If the doodad is non-learning, it may be that it is "pre-trained" before instalation - and yes, statistical math is used for the recognition bit. Vector math is the most common framework. I did a project once where some banks used a NN doodad to detect fraudulent loan applications - lifted the detection rate from 70% to 99% ..

My recent robot session worked well when I shouted with enough invective to cause constant clipping in the audio signal (think Marshal amp with all the knobs turned to 11) - I swear the plastic on the curly handset cable was getting hot!

Mind you - I had plenty of motive:

In their infinite wisdom, my government has decided that all ISPs in Australia should start filtering the internet to prevent bad-lichen and criminal stuff getting across the border. Very laudable, however, very stupid. They have an initial blacklist that got leaked to wikileaks (google wikileaks Australian Government blacklist) .. one or 2 ISPs are now conducting a trial for the filtering.

Unfortunately, my websites and email are on sub-domains/add-on domains with the result that they share a root IP address with a couple hundred other sites on the host server. This has the result that if only one of the hundreds of sites on that domain gets on the blacklist then the whole lot get denied access (null routed at the ISP).

When I found my sites and email were not working I got entangled with the voice robot and had 3 glorious days of chatting with nice Indian girls from my ISP and surly techs from my Host. After the 3 days of running trace routs and sending them hither and thither and getting all my friends to do traces and compiling triangulation sets ... after all that, my ISP fessed-up to having my domain blocked for "active phishing". Well, I am intimately familliar with all the code on my websites - they be squeeky clean on account of my previous host let a hacker in and caused me to know what malicious code looks like!

Once I knew it was there, it took me .. hmm .. 20 minutes to circumvent their block and now I am consulting with the national regulator to see if I can get paid for all the tracing work I did for the ISP plus loss of business for the 3 days they took to admit they were blocking my sites plus $30 for the cost of applying a new root IP.

At the end of the day - the internet is global jurisdiction - it cannot be legislated at a local level. And TCPIP is, after all, a military protocol that was designed to be impossible to interupt. If my government moved to solve the jurisdictional issue then I could support their initiative, but dumbness does not inspire my next vote..

I can see where my government's little exercise in futility will soon result in a lot of angry voters. Perhaps in time the use of subdomains for hosting will be abandoned - but the cost will go out to the community at large with a lot of inocent web users being penalised for the actions of the few. Guilty-until-proven-inocent is not the rule-of-law and can be reliably used to identify a corrupt government in every case.

For myself - I am not a network tech - I have to ask those of you who are: Is there any way to deny access to sub-domains without blocking the root IP?

I will need to know these things as I garner community resistance to the government's faulty legislation. And - this will also be an issue in your country when it gets around - it is wonderfully effective window dressing that any representative group can use to blow taxes into their friends' pockets and keep everyone thinking they give a damn. BullS**T Baffles Brains.

[MTGuru]

Just to disentangle the confusion ... Not all learning systems are neural networks. Statistical systems, such as Bayesian predictive systems, may also require training, but based on very different non-neural models. As Director of Linguistics for one of the world's largest developers of voice recognition technology, I knew something about this stuff, though VR was not my area of involvement.

[Mitch]

hmmm quite right - I have not been following the things closely for a few years. Although, I have used the bayesian system - it seemed to be doing the same stuff as NN from a different angle. Statistical predictive systems need history to function. All this stuff gets into the bowels of suppy chain management software and I've forgotten a lot since I did those things for retailers and such. So I'd presume that any statistical doodad would have some pre-arranged "history" stuck in it for a stand-up-start .. presumably the poor testers would have been subjected to the same stresses while recording the "history" and end up in shouting matches just like a real customer might.

Personally I like the NN gizmos - you can dump them in arrays and make the synaptic loadings show as colors in a frame. Watching them learn is like watching snowflakes - wonderful stuff!

It's all whimsy really - regardless of the doodad behind the pleasant-sounding-but-annoying fembot on the phone - I find that MR Agro gets results. AND I like shouting - so it has become one of life's little pleasures .. it's all good

Incidentally - MT - do you know if there's more surgical ways to block an offensive website without null-routing entire communities?

[Coffee]

Is there any way to deny access to sub-domains without blocking the root IP?

This is exactly the kind of question I should have a ready answer for. Unfortunately I'm drawing a blank at the moment.
I think that it should be doable on a LAN but would be exceedingly difficult to implement on a national level.

[Mitch]

Yes - it's going to be a real big issue - anyone who can get an answer for it will have access to billions of tax dollars around the globe.

The thing is that the internet is a very new planet - it is barely explored. We have all found priceless resource within it, but so too have the exploiters and predators.

Our respective govenments have our mandate (and consensual taxes) to fill the power niches that affect us on our behalf - King for hire (democracy). Their main job is to stop other kingdoms getting control of our territory.

They do this by recognising all power systems within their borders and gaining control over them - commerce, common law, keeping the peace etc - but there is so much more needed to properly fill power niches - so other organisations that do all the stuff that is not "nice" (prostitution, drugs, contract killing, revenge etc) these all get controled secretly by our governments on our behalf - the alternative is that our elected governments get overthrown by the not-nice powers (see Bolivia, etc).

Problem with the internet is that there are no borders - so planet internet is essentially lawless until the world appoints some global authority to occupy it.

For now that leaves us with only 2 alternatives to stop predators raping everything:

1.
Block them at the door (parental control doodads and such voluntary measures) - obviously this must fail because people are mostly good-hearted, but dumb, sheep that do what they are told - and don't do anything unless they are told - mostly they have their ears shut and can't be told, and mostly would not know a pearl from a Mentos mint and would not believe it if you rubbed their snout in a heap of diamonds.

2.
Empower a global consentual body to be the internet police - these would have the power to not only shut down criminal websites and internet nodes, but to publicly dissembowel the instigators on youtube (the law must be SEEN to be observed).

Option #2 is the obvious success path - however, it is unlikely to happen because no power-occupant will cede to a higher authority (they all secretly dream to become the higher power and they will certainly not allow ANYONE to get there before them!).

There is hope in the UN - but the USA systmatically de-boned that hope in Iraq and Guantanamo - I give it a hundred years to get going again .. maybe after humanity gets culled through global warming? The grim reaper is still our shepherd - if humans were worth a cent more than any other animal (I include slime-mould in this definition) then they might have earned a seat above the reaper. Not before.

But still - there is created a void - either the "Russian Mafia" <woooooo -- booger booger booger> will take on the mantle, OR, the gestault of internet standards associations will take the throne. After that, we as punters, need only pay the protection to our overlord or be content with our fringy seat on the edge of the cliff .. like desperate demented seabirds.. When the cull comes - few will be prepared.

What is left for the wise is to begin researching the highways and byways for the new police - better to be a villain's whip than a martyre's skin.

Most the high priests of technology are autistic - it is no mistake that autism is an epidemic. And the high priests sit on all the critical levers - be one or know one. Their levers are obvious if you know where to look.