neural network drum track composition dan smith. goal develop a neural net which can be trained to...

15
Neural Network Drum Track Composition Dan Smith

Upload: kory-burke

Post on 27-Dec-2015

213 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Neural Network Drum Track Composition

Dan Smith

Page 2: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Goal

• Develop a neural net which can be trained to produce drum tracks given a few starting beats.

Page 3: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Previous work

• Michael Mozer – Neural Network Music composition by prediction

• Adam Guetz and Tony Lee – Neural Network Music Composition

Page 4: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Network

• Each note is presented to network, the network predicts next note.

• Training mode– At each note, error is calculated based on next

note– Weights are updated

• Simulation mode– Next note is fed back into network

Page 5: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Network architecture

Next note

Context

Note selector

Neural Network

Page 6: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Note representation

• Note is respresented at duration and instruments

• Six instruments: snare, tom1, tom2, cymbal, hi-hat, bass.

Page 7: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Instrument Representation

• Instruments are represented in binary, 1 means the instrument is being played, 0 means it is not

• Vector is (snare, tom1, tom2, cymbal, hi-hat, bass)

Page 8: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Duration representation

• Time is divided into ticks. Each tick represents 1/12 of a beat. – Eighth note = 6 ticks– Eighth note triplet = 4 ticks

• Note is represented by elapsed time, as well as (elapsed time mod 4) and (elapsed time mod 3)

Page 9: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Why?

• Eighth notes = sixteenth notes mod 3– 3 mod 3 = 0– 6 mod 3 = 0

• Eight note triplets = quarter note triplets mod 4– 4 mod 4 = 0– 8 mod 4 = 0

Page 10: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Duration(cont)

• Each mod is represented by a 1 hot code– Eg 6 mod 4 = 2 = 0100– 6 mod 3 = 0 = 001

Page 11: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Total representation

• 14 until vector

• (duration, mod4 (4 inputs), mod3 (3 inputs), snare, tom1, tom2, cymbal, hi-bat, bass)

Page 12: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Converting Output to Next Note

• Duration is distributed, must pick real duration

• Pick duration vector closest to the duration produced by the network

Page 13: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Results

• Network learned simple patterns without repeated note (analogous to previous work)

• Network learned patterns with repeated notes

• Did not generalize well

Page 14: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

Future suggestions

• Work on generalization– Problem with context resetting

• Try top down production Method

Page 15: Neural Network Drum Track Composition Dan Smith. Goal Develop a neural net which can be trained to produce drum tracks given a few starting beats

References

• Michael Mozer – Neural network music composition by prediction– http://www.cs.colorado.edu/~mozer/papers/

music.html

• Tony Lee and Adam Guetz – Neural Network Music compostion – http://www3.hmc.edu/~anlee/cs152/