On January 4, 2011, Bezos sent an email to a few select
executives including his technical advisor or TA, Greg Hart, “We should build a
$20 device with its brains in the cloud that’s completely controlled by your
voice.” By then Hart and Bezos would have perhaps discussed the advent of
speech recognition multiple times including once in late 2010 when Hart demonstrated
Google’s voice search on Android to Bezos. Soon after the email, Hart was
assigned the secret project to build such a device and he moved out from his TA
role.
By early 2013 the product codenamed Doppler was in beta
testing mostly from a few hundred Amazon employees. The data was noisy and it
wasn’t enough. The device didn’t look promising yet. Bezos kept asking, “How
will we even know when the product is good?” Rohit Prasad was leading the
effort in building a far-field speech recognition engine, a core part of
Doppler. Prasad and Hart prepared a graph showing how Alexa would improve as
data collection progressed. For each successive 3 percent increase in accuracy, they needed to double the data.
In one of the reviews with Bezos, Hart, Prasad, and team
proposed to double the speech science team and postpone the launch from summer
to fall. Bezos commented, “You are going about this the wrong way. First, tell
me what would be a magical product, then tell me how to get there.” Then the
question came whether the team had enough data. Prasad answered that they need
thousands of more hours of complex, far-field voice commands.
Bezos factored in the team’s request for additional speech
scientists and calculated how long the revised team would need to get the
requisite data. He asked, “Let me get this straight. You are telling me that
for your big request to make this product successful, instead of it taking
forty years, it will only take us twenty years?” His math was correct and the
team didn’t have a plan. Bezos ended the meeting abruptly say, “You guys aren’t
serious about making this product.”
This review resulted in a serious introspect in the team and
they came up with a plan of outsourcing the data collection. The new program
called AMPED ended up renting neighborhoods first in Boston and later
expanding it to 10 cities and recruiting contract workers. The rented homes and
apartments were sprinkled with Alexa devices disguised as pedestal microphones,
Xbox gaming consoles, TVs, and tablets. The contract workers worked eight hours
a day, six days a week, reading scripts from iPad with canned lines and open-ended requests.
By 2014, the Alexa team had increased its store of speech
data by a factor of ten thousand. When all this was presented to Bezos in the
review, he said, “Now I know you are serious about it! What are we going to do
next?”
No comments:
Post a Comment