AI for Indian Music

[JLC JLC]

Have a question - Google Magenta will machine learn only from musical notes of song or from the frequencies too

Because Indian Music is highly based on the microtones (which we call it as Gamagam - Change to frequencies in between musical notes)

Please educate.  10 years back I used to write algorithms to generate carnatic based music (Using rules not by AI)

[M4 speers]

Based on the evidence I've heard,  it seems like the AI is taking in all of the raw frequencies and then developing its own harmonic "rules" to recreate the original source.  Please community forgive me if I'm wrong,  but I think I'm hearing whole frequencies manipulation and synthesis rather than just using the "rules" of note relationship which are culturally subjective.  

The MLK speech sounded like music,  but (I'm just listener, and might be wrong) I thought it was from the natural physics of sound waves and interactions between/among them,  not from the AI looking at speech as a system of notes in any established cultural tradition.

I can't get my answer into one succinct sentence, but what do you think?

--
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.

[Paul Cohen]

There are also many other music genres that are texture, layer, and
frequency based.
It's an interesting question. As a slide player myself I'm all over the
frequency scale when I play.

On Tue, 8 May 2018 at 19:29, M4 speers <seagul...@gmail.com> wrote:

> Based on the evidence I've heard, it seems like the AI is taking in all
of the raw frequencies and then developing its own harmonic "rules" to
recreate the original source. Please community forgive me if I'm wrong,
but I think I'm hearing whole frequencies manipulation and synthesis
rather than just using the "rules" of note relationship which are
culturally subjective.

> The MLK speech sounded like music, but (I'm just listener, and might be
wrong) I thought it was from the natural physics of sound waves and
interactions between/among them, not from the AI looking at speech as a
system of notes in any established cultural tradition.

> I can't get my answer into one succinct sentence, but what do you think?



> On Tue, May 8, 2018 at 5:11 AM, JLC JLC <xarun...@gmail.com> wrote:

>> Have a question - Google Magenta will machine learn only from musical
notes of song or from the frequencies too

>> Because Indian Music is highly based on the microtones (which we call it
as Gamagam - Change to frequencies in between musical notes)

>> Please educate. 10 years back I used to write algorithms to generate
carnatic based music (Using rules not by AI)

>> --
>> 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.

> --
> 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.

[Ravi Annaswamy]

Great question. A MIDI note-based system can only from notes, but ML/Magenta can be extended to learn from the raw frequencies (pitch lines or f0 fundamentals)

or even from the raw audio.

The onsets and frames software which is part of magenta extracts Piano notes because it is trained on piano music and so it 'quantizes' the pitches to a piano keyboard notes.

**

Regarding Carnatic gamagams the most breathtaking rule based implementation is Mr.Subramanians Gaayaka software which is freely available to try,

I came across it only a couple of weeks ago. (Pardon me if YOU ARE Mr.Subramanian)

rasika1.apps.com

As you have correctly said, Indian classical music when it chooses a scale, has a specific tuning (non-equitemperamental microtones, e.g, anandha bhairavi's

Ga and Shiva Ranjani's Ga may be different though a key board player is forced to used the same note for both)

In addition to the microtone, the slides and ornamentations are also selective not to disturb the mood.

What I was so amazed was Mr.Subramanian's excellent study of it using spectrograms and his software that can take a melody outline and embellish it into

a slide-y flute or veena rendering.

**

With Machine learning it is possible to learn such a conversion routine. Here is how we would go about it:

1. From audio rendering extract f0 (using excellent continuos pitch extraction algorithms

2. Label the discrete note outlines as in SRGM notation.

3. Use supervised learning to learn the mapping from the note outlines to pitch outlines.

Definitely doable, if someone has the time.

**

With such population, such talent in music, such rich long tradition, it is a shame that the country cannot yet recognize, support and build upon such

rare work as Rasika/Gaayaka software. Hope the time comes soon.

Ravi

[Ravi Annaswamy]

I recognized that Mr.M.Subramaniam spells gamakam and you spell it gamagam, so you are not him :)

Gamaga is a nice onamotapoia word, but gamaka makes it into a verb I think :)

Ravi

[Ravi Annaswamy]

Upon some more thinking, I think that step 2, manual labeling is not needed.

A quantization of the pitches into the SRGM Arohana/Avarohanam should give a SRGM notation of the retrieved pitchline

and an ML sequence to sequence algorithm could learn to embellish it back to the pitchline. Just my guess.

On Tuesday, May 8, 2018 at 8:51:03 PM UTC-4, Ravi Annaswamy wrote: