TuneStar 0.1b Music Generator (feedback appreciated)

[Alejandro Ruiz]

https://soundcloud.com/user-610922241

Hi, can you help me improve my program to generate better music? Also I have 2 different styles I'm working with, one is tunes #1-#4, the other is #5-#7. I would appreciate some feedback on which one sounds better to you.

Thanks!

[Curtis "Fjord" Hawthorne]

Hi Alejandro,

Sounds great! Can you give some more details on how your generative program works?

-Fjord

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discu...@tensorflow.org
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.

[Alejandro Ruiz]

Repetition is planned to some degree. Mainly 4 bars sections with some duration variations. The rest is random probabilities of note duration, pitch, chords, note density and whatnot. A lot of it can be semi-adjusted with parameters but it's all chance. The probabilities are taken from 110 midi songs that I selected.

[Alejandro Ruiz]

The output sounds interesting to me. Much more than the "noise" I hear a lot that most people make when they attempt to make a computer compose for them. And I don't think I'm casually imitating anything, at least the output doesnt plagiarize the corpus I gave it.

[Alejandro Ruiz]

Maybe is because I dont speak english as my first language but I dont understand much of what you are saying. Can you help me improve my program or give me some directions in a clear way please?

[bob Desnos]

where is the app? for mac or win ?

[bob Desnos]

I have a great project in mind

between psychology, art, science

--

Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org

To unsubscribe from this group, send email to magenta-discuss+unsubscribe@tensorflow.org

---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

--

Mauro. M.

A.K.A. 

Bob Desnos

"Cambia il tuo modo di guardare le cose, e il mondo e le cose si trasformeranno "

[Alejandro Ruiz]

There's no app yet, I'm just working on the project. Did you listened to the tracks? Did you prefer fast or slow ones?

[Ravi Annaswamy]

The fast style is better, probably because the melodies are shorter and easily graspable.

[Alejandro Ruiz]

Thanks for your feedback, I think the same but I wasn't sure.

[giancarlo iannizzotto]

I found the faster very nice and easy listening. The style is pretty much similar to the Cure, but without any hint of plagiarism. Overall, I like it but I would better understand how did you make your program learn the probabilities and how much it can generalize or further learn from samples in order to generate other styles of music. Did you write a paper?

[giancarlo iannizzotto]

While listening the slower ones, I got a feeling of Sisters of Mercy or similar dark groups from the late eigthies. Again, pretty good music. The same comments as above hold.
Thank you for sharing your results, and please let us know more about your research.

[Paul Cohen]

Compositionally I thought they were interesting, but from a musical performance perspective too robotic.

By that, I mean missing dynamics and emotional feel. This is a hard problem though, so a good start.

[Alejandro Ruiz]

Thanks Giancarlo, It's funny that I thought of The Cure too when I first listened to the fast tunes. Even when I didn't planned to mimic their style, I think that happened because I made the program use the relative minor key on verses to make the choruses stand more, which gave the tunes a more darker feel. Same with the slow ones. It's also interesting that there was no Cure song in the dataset I used. About your questions, the probabilities come from a Markov chain of all the songs. But I think what makes the program work well is all the rules I constrained the output to comply. Rules like which chords can it use at the beginning and end of verse and choruses. What range the melody can span and more... I haven't wrote a paper yet because I still want to improve it a lot, but probably will do when I'm done. 

Thanks Paul too, the performance is not the priority at the moment. But it can be humanized in the future.

P.S: I added 2 new tunes with different rules than before, please let me know what you think about them.

[Oscar Friedman]

Hi Alejandro,

I am curious why the melody instrument always starts on beat one. One of the unspoken rules of modern pop music is that if the chorus starts on beat one, the verse must start an eighth note before or after beat one. This rule is adhered to around 90% of the time so it's surprising to me that the model didn't pick up on that.

Cheers,

Oscar

--

Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discuss+unsubscribe@tensorflow.org
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

--

Oscar

[Alejandro Ruiz]

Hi Oscar, yeah I saw that too. Remember I'm not using a neural network here, it's basically a Markov chain but with lots of constraint rules on the output. So the model is as smart as the rules I impose it to comply. The way it's encoded right now it's pretty hard to impose that kind of rule, but it can be done maybe in the future. For now it sometimes randomly starts the chorus melody not in beat 1 which is nice because it always does that on the verses.

P.S: Tell me, if you can, which tunes do you think are best to your ears? It helps me know in which directions to move.

[Oscar Friedman]

Hi Alejandro,

I see. Tune number 7 is the most "reasonable" from a compositional standpoint. The verse, pre, and chorus are distinct from eachother. A bit like a Metric song or something. I would also add a rule so that the melody is allowed to take a rest of a quarter or half note at least during the pre. Yours fill every beat of the song which is not so common.

Oscar

On Sun, Apr 8, 2018 at 4:45 PM, Alejandro Ruiz <arodr...@gmail.com> wrote:

Hi Oscar, yeah I saw that too. Remember I'm not using a neural network here, it's basically a Markov chain but with lots of constraint rules on the output. So the model is as smart as the rules I impose it to comply. The way it's encoded right now it's pretty hard to impose that kind of rule, but it can be done maybe in the future. For now it sometimes randomly starts the chorus melody not in beat 1 which is nice because it always does that on the verses.

P.S: Tell me, if you can, which tunes do you think are best to your ears? It helps me know in which directions to move.

--

Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discuss+unsubscribe@tensorflow.org
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

--

Oscar

[Alejandro Ruiz]

Yeah, I can't at the moment figure out a way for making the melody not use every beat. I tried it thought but the results were not good. It's like something is missing. Never heard of Metric, I saw them on youtube but I don't see the resemblance yet. Thanks for your feedback!