I just phoned Aviva about my car insurance. Man, that was a barrel of 
laughs!

First of all, you have to sit through a five minute message telling you 
how much easier it would be to do whatever it is you're trying to do 
over the Internet. [Which I already tried, by the way. There doesn't 
appear to be a way to do what I want to do.]

Five minutes later, we get to the part where they actually "welcome to 
Aviva", and get to handling my call. Now usually this is where you get 
read a list of options, and asked to press a number. ("For Animal from 
The Puppets, press ten!") However, this particular call platform wants 
me to *say* what I want.

Summoning all my years of software engineering and AI design experience, 
I take a flying guess at which combination of keywords might cause the 
machine to route my call to the correct department. (Note that Aviva do 
insurance for just about everything - life insurance, house insurance, 
travel insurance, pet insurance... So it's kind of critical that I get 
the correct department.)

Astonishingly, the machine managed to correctly parse my speech. (I 
didn't think about this at the time, but the voice is recorded rather 
than synthesized, so they must have recordings for the exact words I 
uttered.)

Then it wants my policy number. Which is amusing, because there's 
letters in it. Usually they either get a human to ask for it, or else 
you dial it on your phone - but the latter won't work for letters. 
Instead, they want me to *say* the policy number.

By the way: kudos to whoever thought it was a good idea for a policy 
number to start with the letter O followed by three 0 digits! :-P If the 
recording hadn't told me the first character is always a letter, I would 
have never known. There's almost no distinction in the typeface they 
used to print my policy document...

Anyway, I utter the various letters and digits, and in a spectacular 
feat of digital signal processing, the machine actually understood what 
I said correctly. At this point I'm astounded. [Then again, I did say it 
in the tone of voice you might talk to a toddler with.]

Next it wants to cross-check. So it asks for my sirname. Now this is 
entertaining. Given that the relationship between the characters making 
up a name and the actual way you pronounce it is vague at best (Mrs. 
Bucket? It's bouquet!), I'm left wondering how the hell the software 
actually manages this cross-check. Still, it didn't complain.

Then it asks for my date of birth. Stop and think about this for a 
moment: 1980. While most people would read this as nineteen eighty, it's 
also quite valid to utter one thousand, nine hundred and eighty, one 
nine eight zero, or even just eighty. The software presumably has to be 
able to handle all of these. Also, one might utter the month as March, 
three, ow-three, zero-three, or third. Again, the software seemed happy 
with what I said.

Finally, at long last, it says "I'm transferring you now".

This is immediately followed by "we're helping a lot of customers today. 
I'm sorry about this." Uh... I'm not sure that's exactly what you meant, 
eh? ;-)

Then I get the whole "you know, you could do this online much quicker". 
Again.

Then they say "we would like to offer you a choice of what music you 
listen to while you wait". FOR REAL??! OMGWTFBBQ! "We're put together a 
selection of different types of music." Oh God, they're not going to ask 
me to *say* a selection and try to guess which canned music matches what 
I uttered, are they?? "So, press 1 for some traditional jazz, press 2 
for classical, pr-" I pressed 2.

Oh, what a surprise. A relaxing piano concerto is too quiet to trigger 
the compander on the mobile phone network gateway. Gee, couldn't have 
seen that one coming. So what I get is a lot of stuttering and 
drop-outs. (For those who don't know - and apparently a lot of people 
don't - the purpose of hold music is to reassure you that you are still 
actually connected. If there was just silence, you might think the link 
had dropped.) Also, apparently data compression codecs designed for the 
human voice don't compress piano very well. It sounds very gurgly. 
Presumably the human voice doesn't have so many high partials.

*Finally* I get through to some woman in India. (Why is it always India? 
Why to companies not comprehend that if you're employing somebody FOR 
THE EXPRESSED PURPOSE OF TALKING TO PEOPLE ALL DAY then being able to 
speak their languages is KIND OF CRITICAL?) Anyway, I could just about 
understand what she was saying over the crackly telephone link. 
Eventually I got the quotation I wanted, but they can't give it to me in 
writing. Gee, that's really helpful, right there.

So, a triumph for DSP, and a bit of a failure for customer service...

Post a reply to this message

[snip]
> Astonishingly, the machine managed to correctly parse my speech. (I
> didn't think about this at the time, but the voice is recorded rather
> than synthesized, so they must have recordings for the exact words I
> uttered.)

I work for a very large IT company that does research on language 
recognition, amongst other things.  We were all asked a few years ago to 
spend 15 minutes reciting a long list of words into a voice-mail box. 
This was then used to help their software filter out dialects, accents, 
barytones from sopranos, etc...

-- 
/*Francois Labreque*/#local a=x+y;#local b=x+a;#local c=a+b;#macro P(F//
/*    flabreque    */L)polygon{5,F,F+z,L+z,L,F pigment{rgb 9}}#end union
/*        @        */{P(0,a)P(a,b)P(b,c)P(2*a,2*b)P(2*b,b+c)P(b+c,<2,3>)
/*   gmail.com     */}camera{orthographic location<6,1.25,-6>look_at a }

Post a reply to this message

On 19/05/2011 14:25, Francois Labreque wrote:

> [snip]
>> Astonishingly, the machine managed to correctly parse my speech. (I
>> didn't think about this at the time, but the voice is recorded rather
>> than synthesized, so they must have recordings for the exact words I
>> uttered.)
>
> I work for a very large IT company that does research on language
> recognition, amongst other things. We were all asked a few years ago to
> spend 15 minutes reciting a long list of words into a voice-mail box.
> This was then used to help their software filter out dialects, accents,
> barytones from sopranos, etc...

I actually meant that the reply from the machine is a recorded voice 
rather than a speech synthesizer, so when it says stuff back to me, it 
must have recordings of those words to play.

But yes, I'm sure it takes a long time to tune the speech recognition 
part with different voices and so forth.

Post a reply to this message

On 5/19/2011 3:17, Invisible wrote:
> I'm left wondering how the hell the software actually manages
> this cross-check. Still, it didn't complain.

That's generally an extremely sophisticated algorithm. It likely, for 
example, guesses the nationality of the name and fixes the pronunciation. I 
worked next to the guys doing the voice 411 stuff 20 years ago, and some of 
their stories were pretty fun.

Simple version: Check for plosives, fricatives, and vowels.  Bucket and 
Boquet would be the same. Bucket and smith would not.

> Then it asks for my date of birth. Stop and think about this for a moment:

Again, it's not that hard. There's only a dozen or so ways to say it.

> *Finally* I get through to some woman in India. (Why is it always India? Why
> to companies not comprehend that if you're employing somebody FOR THE
> EXPRESSED PURPOSE OF TALKING TO PEOPLE ALL DAY then being able to speak
> their languages is KIND OF CRITICAL?)

That's why it's India and not China. India was occupied by the british long 
enough that many learn English pretty well.

-- 
Darren New, San Diego CA, USA (PST)
   "Coding without comments is like
    driving without turn signals."

Post a reply to this message

>> I'm left wondering how the hell the software actually manages
>> this cross-check. Still, it didn't complain.
>
> That's generally an extremely sophisticated algorithm. It likely, for
> example, guesses the nationality of the name and fixes the
> pronunciation. I worked next to the guys doing the voice 411 stuff 20
> years ago, and some of their stories were pretty fun.

The English language is such a baroque mixture of incompatible spelling 
conventions, and nowhere does this show up more than in sirnames. I'm 
surprised this can work at all.

> Simple version: Check for plosives, fricatives, and vowels. Bucket and
> Boquet would be the same. Bucket and smith would not.

So it's merely trying to guess whether the correlation looks plausible, 
rather than a definitive test?

>> Then it asks for my date of birth. Stop and think about this for a
>> moment:
>
> Again, it's not that hard. There's only a dozen or so ways to say it.

Well, I might say "March twenty eighth" or "the twenty eighth of March" 
or "the twenty eighth day in March" or several more wordy variations 
than that...

>> *Finally* I get through to some woman in India. (Why is it always
>> India? Why
>> to companies not comprehend that if you're employing somebody FOR THE
>> EXPRESSED PURPOSE OF TALKING TO PEOPLE ALL DAY then being able to speak
>> their languages is KIND OF CRITICAL?)
>
> That's why it's India and not China. India was occupied by the british
> long enough that many learn English pretty well.

Heh. And I thought it was because China is communist...

Post a reply to this message

On 5/19/2011 8:44, Invisible wrote:
> surprised this can work at all.

Well, back when I was looking at it, it was far from a solved problem, but 
there were dozens of really smart people working on it with entire phone 
books worth of examples, so yeah.

Even stuff like "St. Peter St." or "Dr. Martin Luther King Dr." were 
interesting.  Strings of digits would get read out differently depending if 
they were a postal code or a phone number or a street address or a year.

The input side was equally wonky, but I don't remember the details there.

> So it's merely trying to guess whether the correlation looks plausible,
> rather than a definitive test?

I don't know. If I had to do it without a lot of research, that's how I'd do it.

> Well, I might say "March twenty eighth" or "the twenty eighth of March" or
> "the twenty eighth day in March" or several more wordy variations than that...

All of which are pretty easy to put in a list.  Try it.  See how long your 
list is.

> Heh. And I thought it was because China is communist...

I found *very* few people in China speak English. India has enough different 
languages that even in India people tend to speak English a lot.

-- 
Darren New, San Diego CA, USA (PST)
   "Coding without comments is like
    driving without turn signals."

Post a reply to this message

Le 2011/05/19 12:09, Darren New a écrit :

> I found *very* few people in China speak English. India has enough
> different languages that even in India people tend to speak English a lot.
>

Let see, in India there are some 27 different "official national" 
languages, each having several dialects.
Given that you may need an interpret, or two, or tree, to speak to 
someone that alegedly speak your language but use another dialect, it's 
a huge push toward any "common" language.


Alain

Post a reply to this message

Le 2011-05-19 12:09, Darren New a écrit :
>
> I found *very* few people in China speak English. India has enough
> different languages that even in India people tend to speak English a lot.
>
English IS the official language of India.  Most Indians who know a 
second language will know their mother tongue (Punjabi, Bengali, etc...) 
and English.

I have been on conference calls with people from India, where people 
from different regions had to resort to English to speak to each other.

-- 
/*Francois Labreque*/#local a=x+y;#local b=x+a;#local c=a+b;#macro P(F//
/*    flabreque    */L)polygon{5,F,F+z,L+z,L,F pigment{rgb 9}}#end union
/*        @        */{P(0,a)P(a,b)P(b,c)P(2*a,2*b)P(2*b,b+c)P(b+c,<2,3>)
/*   gmail.com     */}camera{orthographic location<6,1.25,-6>look_at a }

Post a reply to this message

On Fri, 20 May 2011 19:41:31 +0200, Francois Labreque  

<fla### [at] videotronca> wrote:


>>
>> I found *very* few people in China speak English. India has enough
>> different languages that even in India people tend to speak English a
  

>> lot.
>>
> English IS the official language of India.  Most Indians who know a  

> second language will know their mother tongue (Punjabi, Bengali, etc..
.)  

> and English.
>
> I have been on conference calls with people from India, where people  

> from different regions had to resort to English to speak to each other
.
>

This is true. I once had a colleague fresh from India. I had to  

continually ask him to repeat slowly over and over until I understood wh
at  

he said. They write much better than they talk.

-- 

-Nekar Xenos-

Post a reply to this message

Am 20.05.2011 22:36, schrieb Nekar Xenos:

> This is true. I once had a colleague fresh from India. I had to
> continually ask him to repeat slowly over and over until I understood
> what he said. They write much better than they talk.

... and remember that when they tilt their head from side to side it's 
not an expression of indecisiveness but their equivalent of nodding.

Post a reply to this message