The Filesharing Conspiracy

THE TEST RESULTS:

The original goal of this test to answer the question of:

1) Can an AUTONOMOUS network of computers using
grid-network-based neural-net software create
an audio music file that resamples an original female
voice into a male voice (Basso Profundo) in a natural
manner such that the NEW audio file SEEMS to normal
everyday human ears and PROFESSIONAL ears to be
both NATURAL SOUNDING and high quality ???

2) A secondary question asked by other posters is:

Can that creation of new high-quality, natural-sounding
multimedia content be made in a timely manner?

---

Those questions can be answered in the following way:

2) In terms of timeliness....after almost 5 weeks and
502 CPU hours later this does SEEM to be a problem for
those who require IMMEDIATE commercial creation
of multimedia content. If you have a low-speed or older
computer system that is single core and less than
1 gigabyte of RAM memory and less than 60 gigabytes
of disk on each system, a 10-node neural network is impractical
for jobs that require a turn-around time of less than one week!

I should add one caveat......in that I did find that our
network administration system tended to KNOCK DOWN
the priority level of the individual grid-nodes which meant
my processsing system did not execute as fast as it could
have, which is why it took almost 5 weeks linear time.
This was the result of the test procedure not being
part of the high-priority "Paying Job" processing queue!

The total aggregate time of 502 hours CPU time is still
there no matter how much linear time was taken
because CPU cycles are a constant no matter what!
My original estimate was quite off-base thus I must
eat crow on this point (hmmm...make that I'll be eating
old & gooey chocolate bars!)

....BUT.... I will say this.....
The ill-timed nature of the test results is thus more
of an administrative problem rather than a technical issue.

On a secondary note, a series of 4 to 10 cheap (i.e. $1400 each
or less!) hexacore AMD computers hooked together on a
high-speed network could make mincemeat out of this task!!!
I would estimate less than one day linear time and 200 hours
aggregate CPU time due to the higher clock cycles and multiple
instruction pipelines inherent with the hexacore systems.


1) NOW THE TEST RESULTS!

The basic fact of the matter is that the neural-net-based
audio-processing system for autonomous high quality
creation of realistic music content is:

NOT YET READY FOR PRIME TIME!

Technical Explanations of Results:

The successes:

a) Specific separation of background instrumentals
and voice tracks was quite successful using brute-force
notch filters. The current limitations on the SIZE of the
rules-of-thumb database and the frequency separation
layers to fit within the available 60 gigabytes on each node
meant each hard disk had to contain a smaller chunk of the
REALLY LARGE 3D acoustic environment map that was
created for processing by individual nodes.

Using a technique similar to Virtual Memory
Page Fault Swapping, individual chunks of the acoustic
environment were swapped out from other nodes to
analyze the paths of the individual "Audio Raytraces"

It is THIS PAGE SWAPPING OVER A NETWORK
which is the main cause for the total length of time
taken by the audio file processing. This can be solved
by installing bigger hard disks and getting more RAM!

Getting back to the point of how a computer can differentiate
between background instrumentals and a voice, there are
tens of thousands of micro rules that use sub-samples
of instruments and short song samples of other material
to compare against a specific audio sample and/or group
of samples. We simply use differential analysis of
acoustic peaks and valley times in specific frequency bands
to ESTIMATE what is considered background music versus
what is vocalization/singing. In computer terms, we XOR
an ESTIMATED background track with the original to separate
out the voice part. We can then TIME the spacing between
words and individual consonants and vowels to get a sense
of overall musical timing. These TIMING SPREADS are then
used to remap and stretch a DIFFERENT VOICE onto the
original voice timings and inflections for individual
word-forms. This is NOT a new technique since it comes
from the science of speech synthesis and voice-print
matching for speech recognition using OPEN-SOURCE
non-patented or expired-patent methodologies.


b) Remapping and pitch-shifting of specific vocalizations
and speech into singing-like intonations: Again, using
grid-based processing using simple brute force DSP
methods is an effective way to match and map external
vocals onto input vocals. THIS IS NOT A NEW TECHNIQUE!

c) Basic acoustic modelling of a 3D physics-based model
of a human vocal tract is an EFFECTIVE and HIGH-QUALITY
METHOD to re-create specific STYLES of intonation and
singing using non-singing spoken words as a source!
THIS WAS THE GREATEST SUCCESS OF THE TEST !!!!
It means I can take spoken words and twist them
into a singing voice!

d) Network bandwidth usage and multi-node grid
processing management has been greatly simplified.
Our software contains a superior node management
system which allows gridding to occur on an
AUTONOMOUS BASIS.

Simple gridding requiring just drag & drop is
a major force-multiplier for use in CONSUMER-LEVEL
processing of multimedia content. That I consider
ALSO a major success for this test. I just let it run
and the computer network neural net management
system did the rest! YOU will also reap the benefit
of that work in OTHER multimedia content creation tasks.

Now for the bad news stated in the post below...



[edit on 2010/8/23 by StargateSG7]

[edit on 2010/8/23 by StargateSG7]

THE TEST RESULTS:

The bad news...

a) Separation of background and foreground and THEN
doing a post filtering to remove reverb/echo introduced
an acoustic artifact called FLANGING which makes an audio
track sound like it was recorded in a small cardboard box.

The main cause is too much high-pass filtering and low-pass
filtering which infringe on the frequency bands inhabited
by the average (er...make that NOT-SO-AVERAGE!) human voice!

The effect is thus caused by a FAULTY RULES-OF-THUMB
database which does not have enough comparisons against
OTHER music and voice type samples to allow a high-quality
differentiation between what is considered an instrument
and a human voice.

HOWEVER this flanging effect can be REDUCED by limiting
the amount of high & lo-pass filtering and doing MORE
Comparisons against vocal and instrument samples
of varying types, intonations and speeds so that differential
frequency-band analysis can be more accurately performed
(i.e. the bad parts are XOR'ed out more often than
the good parts).

Thus I MUST COUNTER that this is an initial rules-of-thumb
initialization and samples database management issue
rather than any technical fault of the Neural Net itself.

b) We hear consonant and Vowel Clipping as individual spoken-word
voice samples are remapped and combined together to
form whole words. In the Youtube video I will be putting
up, you will notice that SOME of the words sound as if they
were simply cut & pasted together to form whole words and
sentences. While the inflections themselves are generally
correct, the sung words sound unnatural because of the simple
fact that actual voice samples were autonomously cut & pasted
together to form whole words, thus you will hear some sharp
transitions between consonants and vowels which give the
singing a robotic feel. I unfortunately am not well-versed
in speech pathology enough to make a PROPER database
of consonant and vowel recombination patterns to allow
natural sounding word-form paste-togethers.

Again, this is a database management issue rather than a
technical fault. To solve this issue there is a requirement for
a SUBJECT MATTER EXPERT (i.e. Speech Pathologist) to form
specific RULES of word-form recombination patterns to allow
for high-accuracy sentence creation in an autonomous system.
Get enough speech pathologists together to agree on a
multi-language word-form re-combination rule-set and
the sharp-sounding transitions will be history!

c) "Pop" introduction as sub-samples are recombined.
You will notice in the final result some "pops & crackles"
which are a result of word-form samples being combined
together which make the audio SLIGHTLY sound like an
old vinyl record. This is caused by extraneous portions
of voice which was not fully filtered out from the
instrumentals track being re-introduced when the
new voice and old instrumentals are re-mixed together.
This is not really a true fault since one could use a
time-based crackle & pop filter to find sharp & short
transients and DSP them out. I'll add that ablity into
the next version. Ergo, this is a programming oversight
on my part! Stupid mistake but fixable!

d) Non-acoustic modelling of actual singing and speech.
To make the voice sound sound less clipped and sharp,
we could ELIMINATE the process of word-form recombination
of many small voice samples by using a truly high-resolution
model of the human larynx to create ACTUAL soundwaves
proprogating through a VIRTUAL VOCAL TRACT!

If someone has the time and money to sample a few hundred
living people of various statures (i.e. physical size) and vocal
range by pumping many sound-waves (i.e. via pink & white
noise generators) through their vocal tracts and using many
tiny microphones to sample the minute echos, reverberations
and sound manipulations performed within that acoustic space,
we could create a database of acoustic propogation patterns
through specific sizes and shapes of human larynxes when
people speak or sing, and then use that acoustic propogation
data to EMULATE A VIRTUAL VOCAL TRACT using physics-based
modelling of a real vocal tract. This would allow us create a
NATURAL speech and singing voice by simply choosing a
desired physical model of a larynx that most closely matches
our desired voice style. The short & sharp staccato
of voice sample recombination would be ELIMINATED
because we could actually SPEAK & SING through a
VIRTUAL VOCAL TRACT that allows virtual sound-waves
of word-forms to propogate through it!

The problem is that SOMEONE has to sample many peoples
physical vocal tracts which is NOT CHEAP to do!
While the sound wave propogation patterns can be described
using Bezier-curve like vectors, those vectors have to be
translated to a final 3D model of X-by-X-by-X number of
physical units of space and time which creates a 3D array
so big that EACH network node machine would need
10 terabyte worth of hard drives and 16 gigabytes of RAM
which is today (as of 2010) still quite expensive!

BUT using that methodology, we could create hundreds of
virtual models of human vocal tracts that have had many
frequencies within multiple audio waveforms PHYSICALLY
SAMPLED AND MODELLED thus creating TRUE speech and
singing voices!

Right now because of storage space requirements, true
physical modelling of multiple vocal tracts is too cost
prohibitive for a CONSUMER ENVIRONMENT!

For commercial users this may not be such a big problem!

See further conclusions below....





[edit on 2010/8/23 by StargateSG7]

THE TEST RESULTS:

The bad news...continued....

e) Bleed-through of old voice tracks.
On the youtube audio sample you will notice
SOME sections sound like there is a second singer
in the background. This is the result of the neural net
not having enough rules to determine which very narrowly
defined frequency bands (i.e. the Notches in a notch filter)
are in the instrumental range or human vocal range!

This is a FIXABLE artifact of not having enough samples
of human voices and instruments to compare against
thus the neural net cannot tell which tiny audio sub-samples
to cut out and which ones to leave as is. Ergo, this means
the old voice is still present on the track that is supposed
to contain only instrumentals or background. Again, this
is fixable by increasing the number of comparison samples.


f) Initial subject matter expert requirements for
rules-of-thumb database creation.

Me doing this test allows me to use my expertise in
audio & video production & post-production as a
gratis (i.e. free!) initial database of initial rules-of-thumb
to embed into the initial neural net.

The problem is that I am NOT a speech pathologist,
nor am I a singer, and nor am I a Bob Ludwig expert
sound mixer. This means that MANY TYPES OF neural nets will
be EXPENSIVE to create because SOMEONE or SOME PEOPLE
who are expert in their fields MUST CREATE an initial set of
rules that make up the decision-making ability of a computer
to define and create high quality multimedia content.

Since their TIME and EXPERTISE is expensive, someone needs
deep enough pockets to PAY for creating those initial
neural net rules-of-thumb databases....At this time, this
constraint won't be solvable by ME!


FINAL CONCLUSIONS:

While I said in an earlier post that this neural net is
NOT YET READY FOR PRIME TIME, I CAN MAKE THIS
DEFINITIVE STATEMENT:

Neural Nets CAN create multimedia content...TODAY that
content is NOT natural sounding or high-enough quality
for professional ears...BUT AGAIN...the system sounds
like someone cut & pasted many audio samples together.
So one could take this tack on the results! It actually did it!
And did it AUTONOMOUSLY! If I had heard nothing but white
noise then I would deem this test a true failure of my software!

However, what I heard was a child doing his/her first cut & paste
and music remix job! The words are intelligible, if low-fidelity
and choppy! This means that with ENOUGH comparison samples
and MORE & BETTER rules-of-thumb the final result will
EVENTUALLY MATCH & EXCEED human abilities!

I will upload the final music video to youtube within a few days.
Because of DCMA issues (i.e. copyright) Youtube tends to remove
Britney Spears songs quickly unless I change them so that
they are considered PARODY or EDUCATIONAL....this means I
am adding a video-based explanation track so that this test
result can be considered FAIR USE and thus be allowed to
remain on Youtube. Don't email me asking for a copy!

Because of the type of work I do, I need to abide by
DCMA/Youtube rules and make some additions to
bring this test result file under FAIR USE territory
and then I can upload it to Youtube for your perusal
and evaluation. Critiques will be welcome and encouraged!


FINAL and SUPER FINAL CONCLUSION:

I think it's funny, you probably will too....it's a work in progress
but the future is pretty bright for eventually using computers
to AUTONOMOUSLY CREATE HIGH-QUALITY MULTIMEDIA CONTENT!

Thank You

and

Happy Multimedia Creation

P.S.

I am updating some of the algorithms to be more discerning
as to what data to examine closer and which to ignore that will
significantly increase processing speed...I will also create
an INITIAL 3D acoustic model of multiple human vocal tracts
by getting a few musically inclined friends to sing and talk
while having their vocal tracts measured using pink & white
noise generators to get sound-wave propogation patterns.

While these 3D models will be rudimentary, they WILL allow
me to create a first start for true vocal tract emulation
software which can be used to TRULY EMULATE human
speech and singing.

Stay tuned for the latest and greatest Neural Net near you...!!!!

P.S.2. Stay tuned for the Youtube test-results link coming soon!

P.S.3. If you have any specific questions, U2U me OR post them
on this thread so that I may answer them for others to read.
I'm a little pressed for time right now because of
work-related constraints, but I will try and answer
your questions in a detailed enough manner
to be truly definitive and clear.

[edit on 2010/8/23 by StargateSG7]

[edit on 2010/8/23 by StargateSG7]

reply to post by Skyfloating


I always download music through filesharing. I have never paid for anything in my itunes library, and I don't intend to either. However, that being said, I think it is sad that everything is digital now. As you described with your book, I like to have something to hold in my hands when I purchase it. When I have a book in my hands I have an actual object, a work of art. The same with a CD or DVD. When my favorite bands come out with new albums I purchase the actual CDs so that I can have the item in my hands and support them in my own little way.

The weirdest thing to me though is the Kindle. Why on earth would I want to download a book and stare at a screen? I spend enough time looking at screens. I like the feel of turning pages, the differing text, etc...I think the Kindle is the most pointless waste of life ever.

So, eh, FAIL?

And I don't agree with your conclusion.

I'm an anarchist, so I am not in favor of legislation to try to prevent people from doing what they are already doing millions of times a day, but in the end, it will just continue to make art a poor profession as no one will want to pay for art or information, since they assume they should get it for free. But after all, who likes a rich, snobby artist anyways? They ought to do it because they love it, not because they want to make money off it. Everything is perfect just the way it is. Artists who couldn't sign big deals can now get their work out to people for free, and the artists who make money make most of their wealth from sponsorship and not customers, so I don't think lady gaga or whatever will be hurting economically if people download her song for free, it will probably just help her out, since now what they are doing is posting a video on youtube for free and running a 30 sec. commercial before you hear the song.

Originally posted by J.Clear
So, eh, FAIL?

And I don't agree with your conclusion.

---


Me thinks the Youtube video will speak for itself.

On a technical note as to the original question
of can a Neural Net make High Quality Multimedia Content?

The basic answer is currently: NO.

However the one caveat I have to add....is that since I didn't
hear "A Cacophonic Splooge of White Noise" I must counter
that YES computers can at least EDIT audio files on
an autonomous basis....BUT...is the result high quality?

I agree in your statement of: *** FAIL ***

....BUT.....Think of what happened....!!!!

a) A professional-quality mainstream-media song was
drag & dropped onto a grid-network icon.

b) Without interference or modification by a human operator,
A song was split apart into chunks, distributed to various
workstations connected together on an el-cheapo network.

Then a series of computer-driven decisions was made to
figure out what the audio waveforms constitute in terms of
what is foreground vocals, what is background instrumentals,
what was the physical size of the original acoustic environment,
what types of special effects or reverberation were added
to the original mix?

A secondary set of decisions was made how to convert a series
of extraneous spoken-word audio waveforms into a format that
closely matches the original inflections, timings and spacing of
a comparison waveform made up of speech-recognized SUNG
words cut out from background audio.

A tertiary set of decisions was made on how to remix the
cut out background with the modified new vocal track
taking into account the original artistic intent of the
old vocals. An output file was created in a desired
musical format playable on many systems without
any other user action or direction.

c) The key point is that I did not have to do ANYTHING other than
create an initial set of database rules-of-thumb on what
constitutes good audio vs. bad audio.

So if you look at it from this perspective, I'll say that while the
test itself FAILED to achieve the STATED GOAL....the fact that
SOME of the sub-goals did in fact get to a certain level of
achievement is a worthy achievement in itself.

I need to learn to crawl before I can walk and then
learn to walk before I can run....It's only a matter
of TIME until YOU will be able to create your own high-quality
autonomously-made multimedia content from user-defined
parameters using an INEXPENSIVE consumer-level neural net!

---
P.S.

I am just adding the explanation video track to the audio
so that Youtube will allow me to keep the test file uploaded
on a more-or-less permanent basis under educational
FAIR USE doctrines. I kinda hate getting cease & desist
letters from copyright holders by uploading technically
copyrighted materials.

So that should be done in a few days. The video portion
will also explain quite fully WHAT was done and HOW it was
done which should give you technical experts something
tasty to chew on in terms of creating YOUR OWN multimedia
content creation software.

Is there a link to the video?

Oh and this thing did not come close to the grand claims you mad. So it failed rather miserably. Bring it back in a decade when you have something near useful.

The original post looks very good. I skimmed it and had to jump in immediately when I saw the references to allowing the population to commit the crime of using illegally obtained files.

With the exception of a very, very few people, the powers that be have something criminal on just about everybody, thanks to file sharing and illegal use of the products of Microsoft and Adobe.

This is serious.

If a government wants to get you for anything and can't make it stick, chances are they can get you on some form of computer piracy. Many, many people could not produce their product license for the Windows operating system itself.

Just a few programs could land your conduct in the category of "grand larceny". That is trouble, real trouble, if some jurisdiction wants to make something of it.

Linux, Open Office or Gimp anyone?

I'd love to see the stats on computer use if YouTube took music videos, except for user/performer generated ones, off their site. How about if copyrighted porn clips disappeared?

There is a lot of public domain content on the web. The web would still be interesting, but it would definitely lose some of it's spice.

[edit on 3-9-2010 by ipsedixit]

I bought this CD awhile ago, for the car. My car doesnt play mp3s. I took the CD up because I wanted to convert it to mp3 so that I could enjoy a certain song on my mp3-player. The CD however, has copyright-protection. So I looked to purchase that single song on the Internet. It was not available for purchase. So I said "what the heck" and downloaded it "illegally". Considering I already paid for the original work, I hereby state that its entirely fine for me to do have done so.

This thread is just about to drop off my radar for watched threads, so can we call this one closed RE:the AI media creation stuff? Weeks of speculation, then weeks of excuses and no final product. Overran the alloted time by aeons and no result.

It was ALL speculation to the max mixed with a big heap of self-promotion and, indeed, advertising. Michael O'Leary/Ryanair would be proud of the tactics used.

4Chan brings MPAA down

The END of Online Piracy

LEAKED EMAILS: Reveals profits of anti-piracy cash scheme

I lost my password for ages So, this is dead, right? Where are the videos??

It has been going on long before torrents or even the internet. I am old enough to remember when Cassette tapes first came out after 8 track tapes. They were the first medium the public could easily copy on a large scale. The music industry whined and cried they would go bankrupt because no one would buy music anymore they would just copy it. But that quickly died down when revenues went up. As it turned out it was the greatest form of free advertising they ever had.

I used to work at a very large well known brand technology company and guys there would buy software and use company equipment to copy the software and even the manual etc. then re shrink-wrap it so it looked like it was never even opened and take it back then everyone else in the lab would get a free copy too if they wanted it. They even ran pirated software son some of the lab servers LOL!

It is the same today. music and video make more money then ever today despite all the downloading. It is free advertising and spreads the word. Most of us know a download is never the quality of store bought so even if I download a good movie or album and really like it I go buy me an original.

I do not think it stifles creativity I think it enhances it the more people are exposed to the more it stimulate creativity. There will always be people with obsessive compulsiveness toward this stuff or anything for that matter it doesn't mean we are regressing.

I read a lot of books I would otherwise never read because I would not spend the money and if good enough I will order a regular copy so I end up buying more books then I would have had I not read the book for free first. Or I will buy several copies and give them away to friends f i think they are that good. Again free advertising and increase of sales.

These industries are making more money then ever and its all because so many people are advertising thier products to friends family etc. for free. The laws against it is not a conspiracy they are part of the control TPTB seek to lock in and this was just another opportunity they see to further their agenda.

Originally posted by Skyfloating


I always read that "downloading them illegally eventually leads to them purchasing it" but I have not seen one single instance where this has happened. Not one.

Haven't seen one? Now you have. I don't download music and stuff. But this applies to me in a different way. In years past when I would read a book from library that I really liked, and knew it was a book I could refer back to for formulas, techniques, recipes, etc. I would order and purchase the book from somewhere. I don't purchase a book unless I have read it first to know it is something I would have a reason to keep.

Google censors BitTorrent, etc.

I welcome myself back to ATS after a considerable absence (8 months or so)
due to a thorough chastising by "those who desire to make lots of money
off my inventions" --- ERGO "Da Big Bosses".

This post is in reply and in regards to my earlier posts indicating that I could
in fact use computer software to modify spoken or sung audio waveforms
to fit any type of acoustic environment and/or voice-print style for the
purposes of relegating media file-sharing to a distant memory!

This type of new technology would in effect preclude the music industry
from suing/gouging/subjugating multimedia content end-users by the way
of allowing powerful software and inexpensive computers to autonomously
create content that would match & mix almost ANY musical or video style
to allow YOU the end-user to create and consume your own self-derived
media content that would MATCH and in some cases EXCEED the technical
and stylistic quality of an original artist.

My attempt to modify a Britney Spears song (i.e. Baby Bay One More Time)
to sound like it was sung by someone else's voice (MINE!) but using the SAME
inflections, rythmic style and tonality was to be demonstrated upon Youtube (R) (TM)
as a expert-systems derived/created recording where computer DSP-like software
would do all the work to convert Britney to Me!

One slight.....eeensy....tiny.....bit....of...a...problem!

Whilst an initial test run WAS IN FACT PERFORMED which at the very least
proved that the concept of voice print re-stylization IS POSSIBLE...a number of
intellectual property and legal issues came up which precluded my initial desire
to demonstrate such techniques in a public forum. (such as on Youtube)

To put it mildly, while my test effort did contain sound quality issues such as
Flanging, Stoccatto Inflection and a few other technical glitches, the underlying
software code technology has a sound basis in principles of Digital Signal
Processing (DSP) and Acoustic Environment Resampling and Reconforming
and is thus EASILY IMPROVED with added funding and technical investigation
to FURTHER REFINE our acoustic re-mapping algorithms.

Ergo, certain honchos who pay the bills found my somewhat-in-jest exercise
to be dubious in nature, unprofessional and on legal shaky ground due to
intellectual property issues which I must unfortunately agree with at this point
in time because I, of course, would ALSO LIKE to be rolling in as much moolah
as possible such that my great, great grandchildren's grandchildren will be able
to uncaringly gouge the peppercorn seeds out of their achingly beautiful,
perfect-chiclet teeth as they wine, dine, and mercilessly taunt the wretched
cement-condo-sucking yuppie huddled masses, yearning to breathe in that
succulent steak seared to sheer perfection within that excessive ostentatious
and HIDEOUSLY EXPENSIVE restaurant which my soon-to-be mucho, mucho
trust fund(s) will be paying for the next 10 generations!

As such, you all get to see ZIP! NADA! NIX, NYET, NAY, NOT A THING, NUTTIN HONEY!
of what I can actually do! Just so you get it it! I'm not taunting you! I'm not doing this
to PO you all off! I'm just doing it out of Gordon Gekko-Greed is Good desires
of eventual riches to be rained down upon my yearning soul by those who wish
to exploit me mercilessly via numerous offers of golden handcuffs and gilded cages!

As the saying goes...IMA SELLIN OUT TO DA MAN CUZ DA MAN
HAS GOT DA GOLD AND I WANT SUM OF IT FER ME AND MINE!

So in conclusion, I must apologize for my long absence and a sincere sorry to those
who have had any unreplied-to emails, but my Yuupie-scum-like pursuit of Benjamin Franklin
has thoroughly corrupted my desire to be an upstanding conspiracy-breaker and
Stick-It-To-Da-Man type do-gooder.

In short, the Grass was Greener on The Other Side...AND IT WON !!! --- FOR NOW!

or for those of you on a more conspiratorial bent....

.....I hereby welcome my new alien overlord masters!