Hi, I’m Kyle.
Hey, I’m Jessie, from the Cornell Lab of ornithology.
This is an experiment we made that uses the machine learning
to organize thousands of bird sounds into the tracked visualization.
Bird sounds can be really difficult to learn because there are so many species out there.
And some of them sound really similar,
and at times you might just need to be distinguishing between the notes that slurs up
or notes that you know drops a pitch, and this can be really really subtle differences.
This project started when we were talking to people at Cornell
about how it might be interesting to play machine learning to bird sounds.
So we decided to try a little test with some of the sounds in a collection.
We didn’t give the computer any outside data or tags, or even tell in the birds names.
We wanted to see if it could just listen to the sounds then automatically learn in its own way to organize them.
Bird sounds vary a lot. So before we can look at them in the same space,
we have to break down everything in short burst of sound less than a second long.
Then we group the sounds for the technique called t-SNE.
首先 电脑创建了一种类似于采集“指纹”的形式 每段声音都如同一种影像
First computer creates a fingerprint form each sound which is like an image
or a set of numbers that represents that particular sample.
Then t-SNE compares all these fingerprints ,and places similar sounds close together.
One way to think about t-SNE
is that it is taking the fingerprints from a high-dimensional space.
That’s spaces were more than three dimensions
and reducing them to two dimensions, so we can visualize them.
And our test worked pretty well. This is a map of sounds that computer created.
You can see that it figured out how to group similar sounds from similar birds close together.
We also made it so that you can search for specific birds.
We put it on the web for you to play with.
当然 我们还有一个远大的梦想 那就是 未来的某天你能够
We certainly have a dream that some day you will be able to put out
a microphone in the Amazon with a recorder and have all the species
you know even not even just birds identified and
we will be able to monitor species diversities outside
through the automated recognition techniques.
I think it’s incredibly exciting to figure out
what those collaborations that we can make between computer science
and ornithology or biology general and start to tackle some really exciting
and challenging science conservation questions.
Visit MacaulayLibrary.org for more resources on a lot of life sounds.
你也可以从 g.co/aiexperiments网站获取源码 参与这项试验 来试试吧~
And you can play with this experiment we made and get code at g.co/aiexperiments.