OK Thanks so much as no matter what he
said the title of my talk is called audio

signal processing with the web or U.P.I.
that's kind of the clunky academical title

it's about what else can we do with
supporting something different which is

not actually all you and if you follow
want to follow along what took.

Cosas lights on your own you
can find them on that link.

And.

That's a little bit of
information about myself

when I was because of quote Undine I'm
basing Golan I'm self-employed and

your coding sets my own
little one man company and

I develop something called Chelsea makes
it's a present based teaching mixer.

And for that teaching mix I wanted to
have an audio fingerprinting system

because there was none in
javascript I had to put my own and

I ported a fingerprint system from
Piscean which was then the reason why is

what about known or
he was sick and possessing and.

Because of my own usual last name I can
use it almost everywhere as a username so

you can find me as could go down the known
as like channel and get up in PM And

while almost all others social networks
out there OK that's is the agenda

that's what I want to talk about it's
what you finish printing in general and

then with a lot of print and
then the actual.

The main part of my talking on all your
second processing and I have to bonus

tracks what I think I won't have time to
go into that in detail and fingerprinting

I think you all know what it is it's the
process of at reading the unique value or

some unique values of an audio signal
which can then be later used to match or

use signals with other or using those.

In a database.

Most popular vacation which is doing that
might be sure Sam but there are of course

open source implementations like because
you could do all of print I could print

pinnacle and I guess there will be
others but I don't know of them and

they all have their advantages at
want to shit and disadvantages and

the end of the day I had to choose one and
I was going to sort of print.

Out of Print is a landmark
based fingerprinting service

fingerprinting system which means it.

Analyzes the frequencies.

And looks in each
frequency pin it looks for

peace once it finds one of those peaks
it compares it to its neighbors and

then it stores the difference of those
peaks and one of the huge loss for

me was that it's used by the Internet
Archive which is one of the audio

sources from a teacher mixes of I haven't
don't have to analyze the files from

the Internet Archive again you can just
use the pix which they've already analyzed

and it's open source in Python which
unfortunately I can't use on the web but

it's a language which I can understand so
it was easier for me to port it and

yeah that was kind of the thing
which I wanted to do a part of it.

Mostly line by line and
at the end I had to work in version but

it was very very slow it was even
slower than the actual audio so

if I wanted to analyze a piece ten seconds
of music it took eleven seconds to analyze

it so it was not the result
I wanted to have and so I.

Came up with a very simplified.

Version of the algorithm which
ought of printers using and

when I was looking at that
to me that looked like

all your tasks even though after
you are applying the F.F.T..

You are not in the ordered world
anymore and you are left alone with

think knows what the tie scsi has to
do with the signals more or less on.

Tars and so my goal was to move as much of
the computation as I could into an offline

audio context and see how that goes and
that's now the main part of my

talk and I want to start by the first
task it was downsampling and

this can be done very easily was to
wear body appealing as might all or

no but there's a big tradeoff because.

The algorithm for downsampling is not
really specified by the A.P.I. And so

you'd probably going to have specified
has implemented its own risen so you.

End up was slightly different versions if
you use down something with a bit more

U.P.I. which then produce likely different
fingerprints if you can live with that you

can use something which looks like that
you are creating and also not a context

the only number the only paramita that
matters is the frequency separate and

then you just decode your only
buffer was that already context and

you get in every buffer at the sample
rate down mixing works very well and

it's specified in the A.P.I. and it's.

Implemented the same in each browsers so
you can safely use it it's a bit

more complicated it looks like that you
create enough unordered context and

then the important part is that
you specify the number of channels

it would have to already Buffalo as
the number of samples and of course

the separate again you have to create
a buffer source node assign your buffer to

the bottom of the buffer source node
connected start it start rendering and

then you get a promise which we
solved to the next audio buffer but

there's a backing from for example you
have to use already buffer which is

already at the decided
sample rate you can't use.

A buffer which has a different sample
rate which would be downsampled but

it doesn't work in grown right now.

Applying the F.F.T. was the next task and
this is very set it does not work

a lot of people tried it and
you can search and stack overflow.

For that.

Many many threads about it and the short
answer is it just doesn't work because

the analyze know to sign for real time and
it will give you the result of whatever

Buffet has currently has and then you
often or your contacts you have no way to

control which whether it's company has so
you can do something like that it looks.

OK but it just won't work so forget it.

But playing the F.F.T..

Was still an important part in my code and
I used an F.F.T.

library at the beginning which was which
was very slow and I came across the bit or

your fifty performance test but.

It's kind of compares different F.F.T.
libraries and

you can see their performance under
different circumstances and for

me that increased my performance
by four hundred percent so

that made me very happy and yet the next
task among this was to find the maximum

in the sick no finding the maximum is not
really all you taught so there's no or

do you know what we can do that natively
but there's descriptors as a nodes and

I thought I could do it with the script
was as a note but after some testing and.

Errors I came across a back which exists
in Chrome Opera and Safari which means

that all you want all your process event
is not called on every buffer so if you or

the Nordic contacts gets large enough you
can almost be sure that there will be

missing process events so
unfortunately you can't use it.

Applying a gate I thought that's a task
for the dynamics compressor node but

it's not supported yet there's a threat
going on in the specification if

it could be used for gate two but right
now you can only use it for compressing

not as an expanded and so I thought I
could use the script was a note again but.

As I already said on or
due process is not called in every.

Event and even worse descriptors as I know
doesn't have any output in all browsers

which very surprised me but
yes you can use it.

The last thing is good news the infinite
impulse response filter is finally

implemented in Chrome and
opera and it works very well.

For me in my test test against my
stupid code it was ten times faster so

it was very words to use it but
unfortunately you can't

reliably reimplement it because of
the broken script process and I would so

you have to use it if it's there and
implement something totally different if

you want to have a solution
which works on all browsers.

So with that that's already my
conclusion these are the good parts

in my opinion so
if you can use the bit or U.P.I. for

computations of arbitrary signals
you have a good chance of.

Resulting in a faster processing time
because it's implemented natively and.

Therefore it should run faster and
of course it's already implemented so

you don't have to do it on your own that
saves you some time and once already work

its will be there which I don't know what
will happen anywhere anytime soon or

the scripts as in those problems are
solved because we don't need it anymore

and even better if suspend remove
suspend resume will get implemented in

all browsers you can use often already
context kind of like a stream when you.

Get Data in chunks you can
process it as a drive to spend

the context wait for the next chunk of
data processing it as planned and so

on and so on and something which
would surprise me is that the web or

U.P.I. can happily handle very large
numbers because I think it's just a float.

Close definition which is the reason for
that and so

the should be capable of for
many many different types of signals.

You know those are the not a good parts.

Can say this program because
of descriptors as annoyed.

As it is already deprecated I don't
think that it will get fixed any more so

there's no way to reliably
polyphase an audio workload

because the script was a No it
doesn't work well that's very bad and

no work or support you can't use off
an already context in a replica which is

country which is normally no problem but
if you have to implement own solutions for

you know which notes which are not SEO yet
you have to shuffle your data back and

forth between the worker and the mains for
that which is not so nice to do.

I don't know how much time do I
have left one woman it's OK so

I can just briefly
mention the bonus tracks.

Or you see able to access the process
project by sub stuck which is mainly known

for building browser if I and he used to
read or U.P.I. to output serial data as

He created or your data which then was
sent to the device and the device into

private data as the real data and
another project called Doppler but

as far as I know he used A.P.I.
to send out high frequencies and

then you could analyze the movement
of the device by that.

Yeah at the end I felt most of
the felt almost like that if I

don't know if you know that it's
a lot instrument I guess and

it says two men with a hammer everything
looks like a nail and for me that would.

Say to a developer with her body VI F.

ing looks like an all you crafts I
don't know if that was the right.

Direction which I took.

Was very interesting journey.

That is like you.

Thank

you.

Thank you I have cheap comments and

one is heck of a tutorial I'd
love to have a copy of that.

And then write in it the second one
is almost on your last slide you

said that you can detect motion by.

Using ultra ultra high frequency yes but
motion.

Motion of the device so
this is not a project of mine so

this is the disclaimer but it works
by sending out high frequencies and

if you move the device the response so
used a microphone to record the response

and if you move to device in response
to change and somehow magically

You can of iced
the movement of the device.

It works I don't know what it was.

Which is already a person with
a word which is audio processing but

it works.

Nothing bad about this no
more questions I can yeah

I was asked to bring up the link for
the selection we have a selection of.

About you croup I don't
know how to say that and

if you want to join selection
you can go to that link and

have a chat with us.

Nothing bad to go early to the demo
sessions thank you very much.

Thank you.