Translation Software That Learns - Is Music Next?

[Garritan]

New Scientist News Service reports about translation software that develops an understanding of languages by scanning through thousands of previously translated documents.

Here's the article: http://www.newscientist.com/article.ns?id=dn7054

The new software, developed by Kevin Knight and Daniel Marcu at the Information Sciences Institute (at the University of Southern California) takes a statistical approach, building probabilistic rules about words, phrases and syntactic structures.

According to the article "The translated documents used to teach the translation algorithms can be electronic, on paper, or even audio files"

Audio files..hmmm. Can this thing learn music? Since music is a language I wonder if we'll be seeing software that learns a person's performance or composing styles and nuances.

Gary Garritan

[Alan Lastufka]

I think all its going to take is for someone/some company to sit an write an exhausting list of rules. That's all it appears this does is "learn" (write a rule) based on what's been done in the past. If a set of rules were to be written for music and notation and sample libs, there would have to be too many standards for all the companies to agree IMHO.

A great idea and I hope it happens soon - would help all of us; Styxx with his theater productions and me with my rock opera, etc....

So Gary, wanna start writing some rules? :D

[lgrohn]

Audio files..hmmm. Can this thing learn music?
Since music is a language I wonder if we'll be seeing software that learns a
person's performance or composing styles and nuances. Gary,

for music the problem might be easier, because music doesn't have semantics
and in a way the syntax is more free. Only the "letters" are much, much more
complicated compared to "normal" languages.

Lauri Gröhn
metacomposer
www.synestesia.com

[Garritan]

Gary,

for music the problem might be easier, because music doesn't have semantics
and in a way the syntax is more free. Only the "letters" are much, much more
complicated compared to "normal" languages.

Lauri Gröhn
metacomposer
www.synestesia.com (http://www.synestesia.com)

I agree Lauri. Music may be easier. This program seems to learn rather than be programmed with rules.

I am wondering if you can plug in every piece of work that say Beethoven, Gershwin or Francesco has ever done - and the program would "learn" their language - their inflections, style, nuances, how they approach phrases, etc. Francesco in a box - what a nice thought.

Gary

[lgrohn]

I am wondering if you can plug in every piece of work that say Beethoven,
Gershwin or Francesco has ever done - and the program would "learn" their language - their
inflections, style, nuances, how they approach phrases, etc. Francesco in a box - what a nice
thought.
Professor David Cope has in a way done this for Bach and some other composers:
http://www.findarticles.com/p/articles/mi_m0FGL/is_1_18/ai_81466098

The CD included in that book gives you an opportunity to quess which piece is composed by human and which by computer.

My own interest is in software that creates original music.. How far can computers go in creativity?
"The problem" is, of course, that we humans still have the listerners' role.... ;)

Lauri Gröhn
www.synestesia.com

[Alan Lastufka]

This program seems to learn rather than be programmed with rules.


...the new software...takes a statistical approach, building probabilistic rules

This will be true for any AI, always. It can't make decisions based on anything other than the rules and laws we give it, whether it is the "three laws" from iRobot or whatever laws this program has - they are always just certain rules.

As much as I wish software could be smart and could understand what I shout at my computer monitor and bang into my keyboard when something isn't working right - it won't happen - until I write a rule/macro that says something to the effect of:

{when "# of keys mashed" = > 1
run -> "Calm_down_Alan_music.exe"}

:D

[SeanHannifin]

Hmmm . . . very interesting . . .

As I've said before in the chat room, if the brain can do it, a computer can do it! (Except maybe be conscious, but how would we ever know that? Or have a soul, if you want to get all religious :D ) I know there has already been a lot of research done with computer composition, and there are some easy algorithms you can use with just the basic rules to come up with some good melodies . . . and a bunch of not so good melodies. However, for the most part, even when you get a computer to create a set of rules based on the properties of other pieces, the end result tends to be gobbledy-gook. Obviously there is something missing from the equation. I'm not talking about emotion. Actually, I don't know what I'm talking about.

Will there ever be a computer that can write good music? I don't know. (Yeah, I know, 'good' is subjective . . . ) Is it possible for a computer to write good music? Of course! But if I ever developed that program that wrote music, I think I'd keep it for myself, and pass off its compositions for my own, and pretend like it never existed . . . if I shared it, all composers would be doomed; if I kept it, the possibilities would be much more grand! Actually, perhaps I'm already using it! :eek: :p

That David Cope book looks interesting, I may have to check it out . . .

[SeanHannifin]

This will be true for any AI, always. It can't make decisions based on anything other than the rules and laws we give it, whether it is the "three laws" from iRobot or whatever laws this program has - they are always just certain rules.

Which is also how the brain composes music . . . :eek: :)

[lgrohn]

This will be true for any AI, always. It can't make decisions based on anything
other than the rules and laws we give it, whether it is the "three laws" from
iRobot or whatever laws this program has - they are always just certain rules.
There are rules on two different levels. The rules of music theory and
the rules of composing. My Synestesia Software has the limited set of rules
of music theory, but after that it takes part of composing rules from any
pictures given to it (based on different rules, of course) and part of the
rules (parameters) from humans. Taking more and more (the rest of ?) parameters
from the picture itself reduces the human interaction. The parameters are re-usable
so in a way one could say that Synestesia Software is an "independent
picture eater composer". Piece number 23 here:
www.synestesia.com
is an example of re-use of parameter. It was composed (the midi file was generated)
in 5 seconds by using parameters of another (not so succesful) piece. But what is best,
now we have GPO and the others making this all much more enjoyable for OUR :) ears, too.

LG

[SeanHannifin]

Lauri Gröhn, I just checked out your website, very interesting stuff! Here I am blathering, and you seem to actually have experience :D

Also, after that secret computer program that writes music is created, another program that does orchestrations would also need to be created :D

[lgrohn]

Also, after that secret computer program that writes music is created, another
program that does orchestrations would also need to be created :D
You are quite right. Until now it has been necessary (for me) to use the least bad
sounds of JV1010. I just used to play the midi files via it without any thought of
articulation etc. Using GPO or the others will be guite a learning process and will
make me, I guess, to change some features my SW. It is a pity that
GPO/Kompact doesn't understand the instrument commands of the midi file.
Or does it?

The process is:

Picture>(select parameters)>generation>midi file>media player> stand alone Kompact >music
or
Select the picture - click the generation - wait 5 seconds - click the midi file - listen great GPO sounds

The SW is not any secret. I have had presentation of it in University of Helsinki
and University of Jyväskylä. The presentation is (so far) only in Finnish:

http://www.kolumbus.fi/lauri.grohn/synestesia.html

[Styxx]

Hmmm, I can see it now ... a computer owning neverland and thus going to trial because it learned all of M.J.'s nuances to play the music perfect! :D

[Markleford]

I'd argue that language has a syntax, but music does not. Language mandates required components and a predictable order to a sentence to maintain comprehension, while music need not be comprehended, only felt. So to my mind (and education), music is *not* a language.

Data correlation and inference are powerful tools in analyzing real world phenomena, but it needs a guiding human hand to know what to look for. Digital signal processing can extract harmonic and rhythmic data from an audio file of music, but how well? Even to this day we don't have a system that can take a music recording and transcribe it to a score.

Without that key technological innovation (namely: translation of ensemble music to data), an unassisted computer will have no more ability to differentiate composing styles than the family dog would.

Generation of music based upon a human-written rule set is something else entirely. I do that myself! But how would a computer ever have the sensitivity to really "get" a performance? It's one thing to analyze rhythmic phrases and convolve new phrases in that vein, but another entirely to display innovation.

Children can string together random phonemes that they've heard: it doesn't mean they speak the language.

- m

[lgrohn]

I'd argue that language has a syntax, but music does
not. If we think syntax as a set of rules which tell which
structures of elemens are right or wrong we could perhaps speak
about syntax of classical music etc. Modern music has less syntax...

So to my mind (and education), music is *not* a language. Perhaps we say that music is language like.
Without that key technological innovation (namely: translation of
ensemble music to data), an unassisted computer will have no more ability to
differentiate composing styles than the family dog would.I am not quite sure what is your point. When Cope's SW can compose Mozart-like
music etc. like music after reading scores, it can in a way
differentiate the styles.
Generation of music based upon a human-written rule set is
something else entirely. I do that myself! But how would a computer ever
have the sensitivity to really "get" a performance? It's one thing to analyze
rhythmic phrases and convolve new phrases in that vein, but another entirely
to display innovation. We all are used to cultural cliches in performances. It may just be
that when listening ultra modern music our cliches are not usable anymore
and we are like Australian aborigines listening classical music?
Children can string together random phonemes that they've heard:
it doesn't mean they speak the language.Not without semantics but if they sign something at the same time...

[Markleford]

I am not quite sure what is your point. When Cope's SW can compose Mozart-like music etc. like music after reading scores, it can in a way differentiate the styles.However, if you feed it half Mozart and half another composer, the computer is going to find commonalities between the two as an amalgamated composer rather than differentiating the styles and throwing the abberant ones out as "noise". The computer will trust your selection of scores 100%: if you say it's Mozart, it will accept the data as Mozart. And that's not even discounting differentiation between early Mozart and late Mozart, or differences between writing for different instrumentation.

My point is primarily that the computer isn't making these inferences on its own: the heuristic is guided by human judgment. And if the human in question does not fully comprehend the dataset, then there's no way he can program a system that can. As such, statements such as "the secret to machine translation is computer power" is completely blue-sky. Feeding infinite quantities of undifferentiated data into such a system is meaningless, as meaning comes from context which is extracted from cognition. In the end, humans are still doing the real work.

That's not to say it's not amazing and worthwhile research. Just that the importance of such research is often overblown by the research organization in question for means of prestige and grant-grabbing, and the press itself often has so little technical ability to discern fact from fiction that most of their reports have a "golly-wowzers!" ring to them.

- m

[Brass Player]

New Scientist News Service reports about translation software that develops an understanding of languages by scanning through thousands of previously translated documents.

Here's the article: http://www.newscientist.com/article.ns?id=dn7054

The new software, developed by Kevin Knight and Daniel Marcu at the Information Sciences Institute (at the University of Southern California) takes a statistical approach, building probabilistic rules about words, phrases and syntactic structures.

According to the article "The translated documents used to teach the translation algorithms can be electronic, on paper, or even audio files"

Audio files..hmmm. Can this thing learn music? Since music is a language I wonder if we'll be seeing software that learns a person's performance or composing styles and nuances.

Gary Garritan
Gary. I'm a new member from the UK and I wondered if you have seen the Tonica Plus software ? This claims to reproduce the performer's style, be that an existing composer or your own. It has what the programmers call a neural network to assess the style/characteristics.

[SeanHannifin]

This Tonica?

http://www.capella-software.com/tonica.htm

It says it will only create the harmony for an already written melody (and they ask quite a price for that).

Neural networks are pretty cool, however, it is nothing new applying them to music. It's been done plenty of times before, and nothing miraculous has yet to come out of the work. The thing neural networks need is input before they can give you any ouput. What kind of input do you give them when writing music? You could give them example pieces, and whether or not you liked the last piece it composed by itself, but beyond that, it's rather difficult. As humans, it is easier for us to compose based on melodies and motifs we know we like. But right now it is impossible for a computer to like or enjoy music . . .

[Brass Player]

This Tonica?

http://www.capella-software.com/tonica.htm

It says it will only create the harmony for an already written melody (and they ask quite a price for that).

Neural networks are pretty cool, however, it is nothing new applying them to music. It's been done plenty of times before, and nothing miraculous has yet to come out of the work. The thing neural networks need is input before they can give you any ouput. What kind of input do you give them when writing music? You could give them example pieces, and whether or not you liked the last piece it composed by itself, but beyond that, it's rather difficult. As humans, it is easier for us to compose based on melodies and motifs we know we like. But right now it is impossible for a computer to like or enjoy music . . .

Thanks for the reply Sean. I fully agree with you. For the time being, I'll sit back and enjoy the sheer pleasure I get from GPO. Hopefully my pleasure will double in the near future with the big band library.

[thesoundsmith]

Band in a Box! :D (Actually, it seems to apply a lot of the same rule-based expert system techniques...)