[ASP]

Hi, folks.

I'm experimenting with MIDI quantization using the Neural Network Toolbox. Any pointers are appreciated.

First, a brief discussion on MIDI quantization. In this specific context its purpose is to "correct" a performance by shifting data representing notes played by an instrumentalist to the nearest subdivision of a beat. More specifically, this scenario involves 16th-beat quantization, which forces the
performance data to shift its positional location to the nearest 1/16th note subdivision within a measure.

My thought is to first create a "template" of sorts. For instance, given a one-measure phrase in Common time (ie, 4/4 time), we subdivide it 16 equal parts, with each part representing a 1/16th beat positional location within the measure. Each part value is the "distance" from the beginning the measure. Conceptually, each value will be a multiple of 1/16th beat in this particular case.

A very basic scenario follows. Given a perforamance of one-measure length in Common time, the goal is to train the network on the 1/16th beat quantization "template", as the template represents the desired target values for the positional location of performance data. Next, the neural network receives as input a one-measure long performance. The task is to examine each note of this performance and shift it to the nearest 1/16th beat for proper 1/16-beat quantization. For example, suppose that a specific note
from the performance falls between two 1/16th beat values. The task is to examine the positional value of that note and then round it off to the nearest 1/16th beat for quantization to occur.

At first I considered this a problem of pattern recognition. For example, in character recognition the character is subdivided into an (m x n) size grid. The character occupies various rows/columns in the grid, and a value of 1 is used to indicate each occupied row/column intersection. These values can then in turn be defined as the training data. The neural network then uses this data to "recognize" characters read into the net that have some degree of noise.

Astractly, my problem is similar to this in that the desired positional data, as specified by the 1/16th-beat quantization template is analogous to the character used to train the network. However, the performance data may likely contain a variety of different beat values as well as rests. For example, in our one-measure Common time scenario suppose the performance data is the following sequence of data:

one quarter note, one quarter rest, 2 1/16th notes, 1 1/8th note and one quarter rest

In one measure of Common time, each of the above elements translate to the following 1/16th-beat equivalents.

one quarter note = 4 1/16th beats
one quarter rest = 4 1/16th beats
2 1/16th notes = 1 1/16th beat each
1 1/8th note = 2 1/16th beats
on quarter rest = 4 1/16th beats

If we sum all the beat values, the result is 16 1/16th beats, which occupies one complete measure in Common time.

This brings us to the problem at hand. How do I approach building a neural network that can examine such input data and produce a 1/16th-beat quantized version. Somehow it will have to read the input data in 1/16th-beat chunks and shift the performance data to the nearest 1/16th-beat value, which is either the most recent 1/16th beat, or the one that follows.

I hope I have described the problem clearly, and I appreciate your feedback.


Thanks.