4

I am a developer but I don't know much about audio and music. I'm trying to detect musical notes in a song file, I'm using p5.sound.js and this script.

I create two arrays 1) Midi values, 2) Start time for that midi value. Then I create a third array for time duration of that midi value.

My draw function looks like following.

function draw() {
  background(200);

  // array of values from -1 to 1
  var timeDomain = fft.waveform(1024, 'float32');
  var corrBuff = autoCorrelate(timeDomain);

  beginShape();
  for (var i = 0; i < corrBuff.length; i++) {
   var w = map(i, 0, corrBuff.length, 0, width);
   var h = map(corrBuff[i], -1, 1, height, 0);
   curveVertex(w, h);
  }
  endShape();

  fill(0);
  text ('Center Clip: ' + centerClipThreshold, 20, 20); 
  line (0, height/2, width, height/2);

  var freq = findFrequency(corrBuff);

  t1_arr.push(source.currentTime());
  midi_arr.push(freqToMidi(freq));

  text ('Fundamental Frequency: ' + freq.toFixed(2), 20, 50); 
}

This code gives me data like this

"midi": [
        "66",
        "74",
        "68",
        "62",
        "69",
        "75",
        "72",
        "66",
        "73",
        "72",....]

"start_time": [
        "2.1066458333333333",
        "2.1119791666666665",
        "2.1386458333333334",
        "2.1546458333333334",
        "2.1653125",
        "2.1866458333333334",
        "2.2079791666666666",
        "2.2559791666666666",
        "2.2826458333333335",
        "2.3093125",....]

"duration": [
        5.3333333333332,
        26.666666666667,
        16,
        10.666666666667,
        21.333333333333,
        21.333333333333,
        48,
        26.666666666667,
        26.666666666666,
        32,....]

The problem is time duration is very short and it also not giving accurate results.

Form this data I'm trying to make tool with a graph like following then I will use mic to record user voice and match it to the notes from the original song and show scores at the end.

Graph screenshot

enter image description here

if there is a better way please suggest me.

James Z
  • 12,209
  • 10
  • 24
  • 44
omi
  • 41
  • 4

1 Answers1

3

Know a bit about music theory, and a bit about audio. Take this with a grain of salt.

What you're asking might be very hard to do. In general, you can have a mix of instruments in a track. Separating each one from the mixed track is known as the cocktail party problem and it doesn't have an easy solution - in fact, solving it is an active area of research.

Even if you simplify the problem so you only have a single instrument, that instrument can play multiple notes at the same time. Think of a chord on a guitar or a typical drum beat.

Finally, if you keep a single instrument playing a single note at a time you might come to something trackable without much effort.

And the FFT will help you here. But you'll need to run it over short periods of time - a small time window. In that window you can look at the spectrum it outputs to try to figure out some dominant frequency that is played by the instrument or singer. But it's going to be noisy. You can look at the dominant in each short window and build a map of the dominant frequency across time. This might begin to resemble something like the midi chart you posted. But it'll still need a lot of massaging to end up with a "note structure". There's also the problem of a particular note not being correctly represented by just a frequency. Might be there's secondary frequencies or phase information you need to take into account.

Further study is needed though. Audio processing with good results is not something that can so simply be done. Ditto for image processing, ML etc.

Horia Coman
  • 8,681
  • 2
  • 23
  • 25