The symbiosis of artificial intelligence and cheap labor
The amazing “ Did everything ” service, produced in 7 days based on some upcoming service for converting voice to text (most likely IVOXMAIL ), impressed everyone with its recognition quality. There are strong assumptions ( in the comments ) that this is not without human operators, because current machine technologies cannot provide recognition quality above 85% (and this is with minimal training under the announcer).
We discussed, argued, and this is what they gave birth (in the words of Alexei Kulakov). What is the most likely IVOXMAIL service:
1. Guys are really too clean recognize the Russian language in order to be able to assume that behind all this is a robot
2. Therefore, we accept that there is a person behind it.
')
3. Next we have the following math. suppose they are counting on a user base of about 50,000 people registered in the first six months. This means that they will have to receive about 10,000 calls per day. Suppose they are not fools and the girls do not answer the call immediately, but simply process the queue from the record to save. This means that they need about 10 girls. Suppose girls are homemade. This means that together with the call center, management and taxes on this business will cost about half a million per month. In principle, this is not a dreadful amount. In addition, this service will be offered to all mobile operators. We consider the hypothesis to be real.
4. What would I do in place of these guys? I would train these neural networks with these girls. Those. I would write a program that breaks the speech into words and matches the text typed by the hands of the girl. and would teach this case neural network. and in a year’s time it would have recognized the Russian language and a popular service for something about 10 million rubles of expenses.
seems to be true?
5. Suppose even that they do not know how to recognize the vocabulary of one person and apply it to another. then it turns out that for each new subscriber they should have a training period, after which the machine learns to recognize the text close to how a girl turns out.
6. the larger the subscriber base, the greater the administrative costs per person and other indirect expenses. Those. each next operator is more expensive. Those. from a certain moment the robot becomes cheaper than a human.
7. 5 and 6 give a model of training for each paid subscriber of the robot by man.
8. If 5 is correct, then it is possible that with the accumulation of statistics a “collective” dictionary will cope with this problem.
Supplement from me:
Most likely, this is arranged using predictive text input, which is substituted by the recognizer. The operator only needs to choose between probable variants or to write completely unrecognized words himself. In this way, both input and training are accelerated at the same time . And in the case when the recognizer works exactly, the operator only needs to press one button.
PS After clarifying about neural networks and voice recognition systems, I will reformulate:
They have not just a neural network there, they have some kind of complicated speech recognition system with neural networks, all sorts of things and dictionaries, which generates predictive text with options for operators, but operators listen to the voice message and very quickly dial / select / accept using this predictive dialing system correct text, teaching the system with its feedback (auxiliary neural networks and dictionaries).
Links that hint at how SPINVOX works (this is a similar bourgeois service):
Comments on Steven Fry's post
Patent
Another anonymous comment about SPINVOX (near the end)
We discussed, argued, and this is what they gave birth (in the words of Alexei Kulakov). What is the most likely IVOXMAIL service:
1. Guys are really too clean recognize the Russian language in order to be able to assume that behind all this is a robot
2. Therefore, we accept that there is a person behind it.
')
3. Next we have the following math. suppose they are counting on a user base of about 50,000 people registered in the first six months. This means that they will have to receive about 10,000 calls per day. Suppose they are not fools and the girls do not answer the call immediately, but simply process the queue from the record to save. This means that they need about 10 girls. Suppose girls are homemade. This means that together with the call center, management and taxes on this business will cost about half a million per month. In principle, this is not a dreadful amount. In addition, this service will be offered to all mobile operators. We consider the hypothesis to be real.
4. What would I do in place of these guys? I would train these neural networks with these girls. Those. I would write a program that breaks the speech into words and matches the text typed by the hands of the girl. and would teach this case neural network. and in a year’s time it would have recognized the Russian language and a popular service for something about 10 million rubles of expenses.
seems to be true?
5. Suppose even that they do not know how to recognize the vocabulary of one person and apply it to another. then it turns out that for each new subscriber they should have a training period, after which the machine learns to recognize the text close to how a girl turns out.
6. the larger the subscriber base, the greater the administrative costs per person and other indirect expenses. Those. each next operator is more expensive. Those. from a certain moment the robot becomes cheaper than a human.
7. 5 and 6 give a model of training for each paid subscriber of the robot by man.
8. If 5 is correct, then it is possible that with the accumulation of statistics a “collective” dictionary will cope with this problem.
Supplement from me:
Most likely, this is arranged using predictive text input, which is substituted by the recognizer. The operator only needs to choose between probable variants or to write completely unrecognized words himself. In this way, both input and training are accelerated at the same time . And in the case when the recognizer works exactly, the operator only needs to press one button.
PS After clarifying about neural networks and voice recognition systems, I will reformulate:
They have not just a neural network there, they have some kind of complicated speech recognition system with neural networks, all sorts of things and dictionaries, which generates predictive text with options for operators, but operators listen to the voice message and very quickly dial / select / accept using this predictive dialing system correct text, teaching the system with its feedback (auxiliary neural networks and dictionaries).
Links that hint at how SPINVOX works (this is a similar bourgeois service):
Comments on Steven Fry's post
Patent
Another anonymous comment about SPINVOX (near the end)
Source: https://habr.com/ru/post/39052/
All Articles